Linux 2.6.21-rc7
[linux/fpc-iii.git] / block / ll_rw_blk.c
blob3de06953ac33d24ce669c35f0e8ffdbb4bcb9661
1 /*
2 * Copyright (C) 1991, 1992 Linus Torvalds
3 * Copyright (C) 1994, Karl Keyte: Added support for disk statistics
4 * Elevator latency, (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
5 * Queue request tables / lock, selectable elevator, Jens Axboe <axboe@suse.de>
6 * kernel-doc documentation started by NeilBrown <neilb@cse.unsw.edu.au> - July2000
7 * bio rewrite, highmem i/o, etc, Jens Axboe <axboe@suse.de> - may 2001
8 */
11 * This handles all read/write requests to block devices
13 #include <linux/kernel.h>
14 #include <linux/module.h>
15 #include <linux/backing-dev.h>
16 #include <linux/bio.h>
17 #include <linux/blkdev.h>
18 #include <linux/highmem.h>
19 #include <linux/mm.h>
20 #include <linux/kernel_stat.h>
21 #include <linux/string.h>
22 #include <linux/init.h>
23 #include <linux/bootmem.h> /* for max_pfn/max_low_pfn */
24 #include <linux/completion.h>
25 #include <linux/slab.h>
26 #include <linux/swap.h>
27 #include <linux/writeback.h>
28 #include <linux/task_io_accounting_ops.h>
29 #include <linux/interrupt.h>
30 #include <linux/cpu.h>
31 #include <linux/blktrace_api.h>
32 #include <linux/fault-inject.h>
35 * for max sense size
37 #include <scsi/scsi_cmnd.h>
39 static void blk_unplug_work(struct work_struct *work);
40 static void blk_unplug_timeout(unsigned long data);
41 static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io);
42 static void init_request_from_bio(struct request *req, struct bio *bio);
43 static int __make_request(request_queue_t *q, struct bio *bio);
44 static struct io_context *current_io_context(gfp_t gfp_flags, int node);
47 * For the allocated request tables
49 static struct kmem_cache *request_cachep;
52 * For queue allocation
54 static struct kmem_cache *requestq_cachep;
57 * For io context allocations
59 static struct kmem_cache *iocontext_cachep;
62 * Controlling structure to kblockd
64 static struct workqueue_struct *kblockd_workqueue;
66 unsigned long blk_max_low_pfn, blk_max_pfn;
68 EXPORT_SYMBOL(blk_max_low_pfn);
69 EXPORT_SYMBOL(blk_max_pfn);
71 static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
73 /* Amount of time in which a process may batch requests */
74 #define BLK_BATCH_TIME (HZ/50UL)
76 /* Number of requests a "batching" process may submit */
77 #define BLK_BATCH_REQ 32
80 * Return the threshold (number of used requests) at which the queue is
81 * considered to be congested. It include a little hysteresis to keep the
82 * context switch rate down.
84 static inline int queue_congestion_on_threshold(struct request_queue *q)
86 return q->nr_congestion_on;
90 * The threshold at which a queue is considered to be uncongested
92 static inline int queue_congestion_off_threshold(struct request_queue *q)
94 return q->nr_congestion_off;
97 static void blk_queue_congestion_threshold(struct request_queue *q)
99 int nr;
101 nr = q->nr_requests - (q->nr_requests / 8) + 1;
102 if (nr > q->nr_requests)
103 nr = q->nr_requests;
104 q->nr_congestion_on = nr;
106 nr = q->nr_requests - (q->nr_requests / 8) - (q->nr_requests / 16) - 1;
107 if (nr < 1)
108 nr = 1;
109 q->nr_congestion_off = nr;
113 * blk_get_backing_dev_info - get the address of a queue's backing_dev_info
114 * @bdev: device
116 * Locates the passed device's request queue and returns the address of its
117 * backing_dev_info
119 * Will return NULL if the request queue cannot be located.
121 struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev)
123 struct backing_dev_info *ret = NULL;
124 request_queue_t *q = bdev_get_queue(bdev);
126 if (q)
127 ret = &q->backing_dev_info;
128 return ret;
130 EXPORT_SYMBOL(blk_get_backing_dev_info);
133 * blk_queue_prep_rq - set a prepare_request function for queue
134 * @q: queue
135 * @pfn: prepare_request function
137 * It's possible for a queue to register a prepare_request callback which
138 * is invoked before the request is handed to the request_fn. The goal of
139 * the function is to prepare a request for I/O, it can be used to build a
140 * cdb from the request data for instance.
143 void blk_queue_prep_rq(request_queue_t *q, prep_rq_fn *pfn)
145 q->prep_rq_fn = pfn;
148 EXPORT_SYMBOL(blk_queue_prep_rq);
151 * blk_queue_merge_bvec - set a merge_bvec function for queue
152 * @q: queue
153 * @mbfn: merge_bvec_fn
155 * Usually queues have static limitations on the max sectors or segments that
156 * we can put in a request. Stacking drivers may have some settings that
157 * are dynamic, and thus we have to query the queue whether it is ok to
158 * add a new bio_vec to a bio at a given offset or not. If the block device
159 * has such limitations, it needs to register a merge_bvec_fn to control
160 * the size of bio's sent to it. Note that a block device *must* allow a
161 * single page to be added to an empty bio. The block device driver may want
162 * to use the bio_split() function to deal with these bio's. By default
163 * no merge_bvec_fn is defined for a queue, and only the fixed limits are
164 * honored.
166 void blk_queue_merge_bvec(request_queue_t *q, merge_bvec_fn *mbfn)
168 q->merge_bvec_fn = mbfn;
171 EXPORT_SYMBOL(blk_queue_merge_bvec);
173 void blk_queue_softirq_done(request_queue_t *q, softirq_done_fn *fn)
175 q->softirq_done_fn = fn;
178 EXPORT_SYMBOL(blk_queue_softirq_done);
181 * blk_queue_make_request - define an alternate make_request function for a device
182 * @q: the request queue for the device to be affected
183 * @mfn: the alternate make_request function
185 * Description:
186 * The normal way for &struct bios to be passed to a device
187 * driver is for them to be collected into requests on a request
188 * queue, and then to allow the device driver to select requests
189 * off that queue when it is ready. This works well for many block
190 * devices. However some block devices (typically virtual devices
191 * such as md or lvm) do not benefit from the processing on the
192 * request queue, and are served best by having the requests passed
193 * directly to them. This can be achieved by providing a function
194 * to blk_queue_make_request().
196 * Caveat:
197 * The driver that does this *must* be able to deal appropriately
198 * with buffers in "highmemory". This can be accomplished by either calling
199 * __bio_kmap_atomic() to get a temporary kernel mapping, or by calling
200 * blk_queue_bounce() to create a buffer in normal memory.
202 void blk_queue_make_request(request_queue_t * q, make_request_fn * mfn)
205 * set defaults
207 q->nr_requests = BLKDEV_MAX_RQ;
208 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
209 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
210 q->make_request_fn = mfn;
211 q->backing_dev_info.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
212 q->backing_dev_info.state = 0;
213 q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
214 blk_queue_max_sectors(q, SAFE_MAX_SECTORS);
215 blk_queue_hardsect_size(q, 512);
216 blk_queue_dma_alignment(q, 511);
217 blk_queue_congestion_threshold(q);
218 q->nr_batching = BLK_BATCH_REQ;
220 q->unplug_thresh = 4; /* hmm */
221 q->unplug_delay = (3 * HZ) / 1000; /* 3 milliseconds */
222 if (q->unplug_delay == 0)
223 q->unplug_delay = 1;
225 INIT_WORK(&q->unplug_work, blk_unplug_work);
227 q->unplug_timer.function = blk_unplug_timeout;
228 q->unplug_timer.data = (unsigned long)q;
231 * by default assume old behaviour and bounce for any highmem page
233 blk_queue_bounce_limit(q, BLK_BOUNCE_HIGH);
236 EXPORT_SYMBOL(blk_queue_make_request);
238 static void rq_init(request_queue_t *q, struct request *rq)
240 INIT_LIST_HEAD(&rq->queuelist);
241 INIT_LIST_HEAD(&rq->donelist);
243 rq->errors = 0;
244 rq->bio = rq->biotail = NULL;
245 INIT_HLIST_NODE(&rq->hash);
246 RB_CLEAR_NODE(&rq->rb_node);
247 rq->ioprio = 0;
248 rq->buffer = NULL;
249 rq->ref_count = 1;
250 rq->q = q;
251 rq->special = NULL;
252 rq->data_len = 0;
253 rq->data = NULL;
254 rq->nr_phys_segments = 0;
255 rq->sense = NULL;
256 rq->end_io = NULL;
257 rq->end_io_data = NULL;
258 rq->completion_data = NULL;
262 * blk_queue_ordered - does this queue support ordered writes
263 * @q: the request queue
264 * @ordered: one of QUEUE_ORDERED_*
265 * @prepare_flush_fn: rq setup helper for cache flush ordered writes
267 * Description:
268 * For journalled file systems, doing ordered writes on a commit
269 * block instead of explicitly doing wait_on_buffer (which is bad
270 * for performance) can be a big win. Block drivers supporting this
271 * feature should call this function and indicate so.
274 int blk_queue_ordered(request_queue_t *q, unsigned ordered,
275 prepare_flush_fn *prepare_flush_fn)
277 if (ordered & (QUEUE_ORDERED_PREFLUSH | QUEUE_ORDERED_POSTFLUSH) &&
278 prepare_flush_fn == NULL) {
279 printk(KERN_ERR "blk_queue_ordered: prepare_flush_fn required\n");
280 return -EINVAL;
283 if (ordered != QUEUE_ORDERED_NONE &&
284 ordered != QUEUE_ORDERED_DRAIN &&
285 ordered != QUEUE_ORDERED_DRAIN_FLUSH &&
286 ordered != QUEUE_ORDERED_DRAIN_FUA &&
287 ordered != QUEUE_ORDERED_TAG &&
288 ordered != QUEUE_ORDERED_TAG_FLUSH &&
289 ordered != QUEUE_ORDERED_TAG_FUA) {
290 printk(KERN_ERR "blk_queue_ordered: bad value %d\n", ordered);
291 return -EINVAL;
294 q->ordered = ordered;
295 q->next_ordered = ordered;
296 q->prepare_flush_fn = prepare_flush_fn;
298 return 0;
301 EXPORT_SYMBOL(blk_queue_ordered);
304 * blk_queue_issue_flush_fn - set function for issuing a flush
305 * @q: the request queue
306 * @iff: the function to be called issuing the flush
308 * Description:
309 * If a driver supports issuing a flush command, the support is notified
310 * to the block layer by defining it through this call.
313 void blk_queue_issue_flush_fn(request_queue_t *q, issue_flush_fn *iff)
315 q->issue_flush_fn = iff;
318 EXPORT_SYMBOL(blk_queue_issue_flush_fn);
321 * Cache flushing for ordered writes handling
323 inline unsigned blk_ordered_cur_seq(request_queue_t *q)
325 if (!q->ordseq)
326 return 0;
327 return 1 << ffz(q->ordseq);
330 unsigned blk_ordered_req_seq(struct request *rq)
332 request_queue_t *q = rq->q;
334 BUG_ON(q->ordseq == 0);
336 if (rq == &q->pre_flush_rq)
337 return QUEUE_ORDSEQ_PREFLUSH;
338 if (rq == &q->bar_rq)
339 return QUEUE_ORDSEQ_BAR;
340 if (rq == &q->post_flush_rq)
341 return QUEUE_ORDSEQ_POSTFLUSH;
343 if ((rq->cmd_flags & REQ_ORDERED_COLOR) ==
344 (q->orig_bar_rq->cmd_flags & REQ_ORDERED_COLOR))
345 return QUEUE_ORDSEQ_DRAIN;
346 else
347 return QUEUE_ORDSEQ_DONE;
350 void blk_ordered_complete_seq(request_queue_t *q, unsigned seq, int error)
352 struct request *rq;
353 int uptodate;
355 if (error && !q->orderr)
356 q->orderr = error;
358 BUG_ON(q->ordseq & seq);
359 q->ordseq |= seq;
361 if (blk_ordered_cur_seq(q) != QUEUE_ORDSEQ_DONE)
362 return;
365 * Okay, sequence complete.
367 rq = q->orig_bar_rq;
368 uptodate = q->orderr ? q->orderr : 1;
370 q->ordseq = 0;
372 end_that_request_first(rq, uptodate, rq->hard_nr_sectors);
373 end_that_request_last(rq, uptodate);
376 static void pre_flush_end_io(struct request *rq, int error)
378 elv_completed_request(rq->q, rq);
379 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_PREFLUSH, error);
382 static void bar_end_io(struct request *rq, int error)
384 elv_completed_request(rq->q, rq);
385 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_BAR, error);
388 static void post_flush_end_io(struct request *rq, int error)
390 elv_completed_request(rq->q, rq);
391 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_POSTFLUSH, error);
394 static void queue_flush(request_queue_t *q, unsigned which)
396 struct request *rq;
397 rq_end_io_fn *end_io;
399 if (which == QUEUE_ORDERED_PREFLUSH) {
400 rq = &q->pre_flush_rq;
401 end_io = pre_flush_end_io;
402 } else {
403 rq = &q->post_flush_rq;
404 end_io = post_flush_end_io;
407 rq->cmd_flags = REQ_HARDBARRIER;
408 rq_init(q, rq);
409 rq->elevator_private = NULL;
410 rq->elevator_private2 = NULL;
411 rq->rq_disk = q->bar_rq.rq_disk;
412 rq->end_io = end_io;
413 q->prepare_flush_fn(q, rq);
415 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
418 static inline struct request *start_ordered(request_queue_t *q,
419 struct request *rq)
421 q->bi_size = 0;
422 q->orderr = 0;
423 q->ordered = q->next_ordered;
424 q->ordseq |= QUEUE_ORDSEQ_STARTED;
427 * Prep proxy barrier request.
429 blkdev_dequeue_request(rq);
430 q->orig_bar_rq = rq;
431 rq = &q->bar_rq;
432 rq->cmd_flags = 0;
433 rq_init(q, rq);
434 if (bio_data_dir(q->orig_bar_rq->bio) == WRITE)
435 rq->cmd_flags |= REQ_RW;
436 rq->cmd_flags |= q->ordered & QUEUE_ORDERED_FUA ? REQ_FUA : 0;
437 rq->elevator_private = NULL;
438 rq->elevator_private2 = NULL;
439 init_request_from_bio(rq, q->orig_bar_rq->bio);
440 rq->end_io = bar_end_io;
443 * Queue ordered sequence. As we stack them at the head, we
444 * need to queue in reverse order. Note that we rely on that
445 * no fs request uses ELEVATOR_INSERT_FRONT and thus no fs
446 * request gets inbetween ordered sequence.
448 if (q->ordered & QUEUE_ORDERED_POSTFLUSH)
449 queue_flush(q, QUEUE_ORDERED_POSTFLUSH);
450 else
451 q->ordseq |= QUEUE_ORDSEQ_POSTFLUSH;
453 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
455 if (q->ordered & QUEUE_ORDERED_PREFLUSH) {
456 queue_flush(q, QUEUE_ORDERED_PREFLUSH);
457 rq = &q->pre_flush_rq;
458 } else
459 q->ordseq |= QUEUE_ORDSEQ_PREFLUSH;
461 if ((q->ordered & QUEUE_ORDERED_TAG) || q->in_flight == 0)
462 q->ordseq |= QUEUE_ORDSEQ_DRAIN;
463 else
464 rq = NULL;
466 return rq;
469 int blk_do_ordered(request_queue_t *q, struct request **rqp)
471 struct request *rq = *rqp;
472 int is_barrier = blk_fs_request(rq) && blk_barrier_rq(rq);
474 if (!q->ordseq) {
475 if (!is_barrier)
476 return 1;
478 if (q->next_ordered != QUEUE_ORDERED_NONE) {
479 *rqp = start_ordered(q, rq);
480 return 1;
481 } else {
483 * This can happen when the queue switches to
484 * ORDERED_NONE while this request is on it.
486 blkdev_dequeue_request(rq);
487 end_that_request_first(rq, -EOPNOTSUPP,
488 rq->hard_nr_sectors);
489 end_that_request_last(rq, -EOPNOTSUPP);
490 *rqp = NULL;
491 return 0;
496 * Ordered sequence in progress
499 /* Special requests are not subject to ordering rules. */
500 if (!blk_fs_request(rq) &&
501 rq != &q->pre_flush_rq && rq != &q->post_flush_rq)
502 return 1;
504 if (q->ordered & QUEUE_ORDERED_TAG) {
505 /* Ordered by tag. Blocking the next barrier is enough. */
506 if (is_barrier && rq != &q->bar_rq)
507 *rqp = NULL;
508 } else {
509 /* Ordered by draining. Wait for turn. */
510 WARN_ON(blk_ordered_req_seq(rq) < blk_ordered_cur_seq(q));
511 if (blk_ordered_req_seq(rq) > blk_ordered_cur_seq(q))
512 *rqp = NULL;
515 return 1;
518 static int flush_dry_bio_endio(struct bio *bio, unsigned int bytes, int error)
520 request_queue_t *q = bio->bi_private;
521 struct bio_vec *bvec;
522 int i;
525 * This is dry run, restore bio_sector and size. We'll finish
526 * this request again with the original bi_end_io after an
527 * error occurs or post flush is complete.
529 q->bi_size += bytes;
531 if (bio->bi_size)
532 return 1;
534 /* Rewind bvec's */
535 bio->bi_idx = 0;
536 bio_for_each_segment(bvec, bio, i) {
537 bvec->bv_len += bvec->bv_offset;
538 bvec->bv_offset = 0;
541 /* Reset bio */
542 set_bit(BIO_UPTODATE, &bio->bi_flags);
543 bio->bi_size = q->bi_size;
544 bio->bi_sector -= (q->bi_size >> 9);
545 q->bi_size = 0;
547 return 0;
550 static int ordered_bio_endio(struct request *rq, struct bio *bio,
551 unsigned int nbytes, int error)
553 request_queue_t *q = rq->q;
554 bio_end_io_t *endio;
555 void *private;
557 if (&q->bar_rq != rq)
558 return 0;
561 * Okay, this is the barrier request in progress, dry finish it.
563 if (error && !q->orderr)
564 q->orderr = error;
566 endio = bio->bi_end_io;
567 private = bio->bi_private;
568 bio->bi_end_io = flush_dry_bio_endio;
569 bio->bi_private = q;
571 bio_endio(bio, nbytes, error);
573 bio->bi_end_io = endio;
574 bio->bi_private = private;
576 return 1;
580 * blk_queue_bounce_limit - set bounce buffer limit for queue
581 * @q: the request queue for the device
582 * @dma_addr: bus address limit
584 * Description:
585 * Different hardware can have different requirements as to what pages
586 * it can do I/O directly to. A low level driver can call
587 * blk_queue_bounce_limit to have lower memory pages allocated as bounce
588 * buffers for doing I/O to pages residing above @page.
590 void blk_queue_bounce_limit(request_queue_t *q, u64 dma_addr)
592 unsigned long bounce_pfn = dma_addr >> PAGE_SHIFT;
593 int dma = 0;
595 q->bounce_gfp = GFP_NOIO;
596 #if BITS_PER_LONG == 64
597 /* Assume anything <= 4GB can be handled by IOMMU.
598 Actually some IOMMUs can handle everything, but I don't
599 know of a way to test this here. */
600 if (bounce_pfn < (min_t(u64,0xffffffff,BLK_BOUNCE_HIGH) >> PAGE_SHIFT))
601 dma = 1;
602 q->bounce_pfn = max_low_pfn;
603 #else
604 if (bounce_pfn < blk_max_low_pfn)
605 dma = 1;
606 q->bounce_pfn = bounce_pfn;
607 #endif
608 if (dma) {
609 init_emergency_isa_pool();
610 q->bounce_gfp = GFP_NOIO | GFP_DMA;
611 q->bounce_pfn = bounce_pfn;
615 EXPORT_SYMBOL(blk_queue_bounce_limit);
618 * blk_queue_max_sectors - set max sectors for a request for this queue
619 * @q: the request queue for the device
620 * @max_sectors: max sectors in the usual 512b unit
622 * Description:
623 * Enables a low level driver to set an upper limit on the size of
624 * received requests.
626 void blk_queue_max_sectors(request_queue_t *q, unsigned int max_sectors)
628 if ((max_sectors << 9) < PAGE_CACHE_SIZE) {
629 max_sectors = 1 << (PAGE_CACHE_SHIFT - 9);
630 printk("%s: set to minimum %d\n", __FUNCTION__, max_sectors);
633 if (BLK_DEF_MAX_SECTORS > max_sectors)
634 q->max_hw_sectors = q->max_sectors = max_sectors;
635 else {
636 q->max_sectors = BLK_DEF_MAX_SECTORS;
637 q->max_hw_sectors = max_sectors;
641 EXPORT_SYMBOL(blk_queue_max_sectors);
644 * blk_queue_max_phys_segments - set max phys segments for a request for this queue
645 * @q: the request queue for the device
646 * @max_segments: max number of segments
648 * Description:
649 * Enables a low level driver to set an upper limit on the number of
650 * physical data segments in a request. This would be the largest sized
651 * scatter list the driver could handle.
653 void blk_queue_max_phys_segments(request_queue_t *q, unsigned short max_segments)
655 if (!max_segments) {
656 max_segments = 1;
657 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
660 q->max_phys_segments = max_segments;
663 EXPORT_SYMBOL(blk_queue_max_phys_segments);
666 * blk_queue_max_hw_segments - set max hw segments for a request for this queue
667 * @q: the request queue for the device
668 * @max_segments: max number of segments
670 * Description:
671 * Enables a low level driver to set an upper limit on the number of
672 * hw data segments in a request. This would be the largest number of
673 * address/length pairs the host adapter can actually give as once
674 * to the device.
676 void blk_queue_max_hw_segments(request_queue_t *q, unsigned short max_segments)
678 if (!max_segments) {
679 max_segments = 1;
680 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
683 q->max_hw_segments = max_segments;
686 EXPORT_SYMBOL(blk_queue_max_hw_segments);
689 * blk_queue_max_segment_size - set max segment size for blk_rq_map_sg
690 * @q: the request queue for the device
691 * @max_size: max size of segment in bytes
693 * Description:
694 * Enables a low level driver to set an upper limit on the size of a
695 * coalesced segment
697 void blk_queue_max_segment_size(request_queue_t *q, unsigned int max_size)
699 if (max_size < PAGE_CACHE_SIZE) {
700 max_size = PAGE_CACHE_SIZE;
701 printk("%s: set to minimum %d\n", __FUNCTION__, max_size);
704 q->max_segment_size = max_size;
707 EXPORT_SYMBOL(blk_queue_max_segment_size);
710 * blk_queue_hardsect_size - set hardware sector size for the queue
711 * @q: the request queue for the device
712 * @size: the hardware sector size, in bytes
714 * Description:
715 * This should typically be set to the lowest possible sector size
716 * that the hardware can operate on (possible without reverting to
717 * even internal read-modify-write operations). Usually the default
718 * of 512 covers most hardware.
720 void blk_queue_hardsect_size(request_queue_t *q, unsigned short size)
722 q->hardsect_size = size;
725 EXPORT_SYMBOL(blk_queue_hardsect_size);
728 * Returns the minimum that is _not_ zero, unless both are zero.
730 #define min_not_zero(l, r) (l == 0) ? r : ((r == 0) ? l : min(l, r))
733 * blk_queue_stack_limits - inherit underlying queue limits for stacked drivers
734 * @t: the stacking driver (top)
735 * @b: the underlying device (bottom)
737 void blk_queue_stack_limits(request_queue_t *t, request_queue_t *b)
739 /* zero is "infinity" */
740 t->max_sectors = min_not_zero(t->max_sectors,b->max_sectors);
741 t->max_hw_sectors = min_not_zero(t->max_hw_sectors,b->max_hw_sectors);
743 t->max_phys_segments = min(t->max_phys_segments,b->max_phys_segments);
744 t->max_hw_segments = min(t->max_hw_segments,b->max_hw_segments);
745 t->max_segment_size = min(t->max_segment_size,b->max_segment_size);
746 t->hardsect_size = max(t->hardsect_size,b->hardsect_size);
747 if (!test_bit(QUEUE_FLAG_CLUSTER, &b->queue_flags))
748 clear_bit(QUEUE_FLAG_CLUSTER, &t->queue_flags);
751 EXPORT_SYMBOL(blk_queue_stack_limits);
754 * blk_queue_segment_boundary - set boundary rules for segment merging
755 * @q: the request queue for the device
756 * @mask: the memory boundary mask
758 void blk_queue_segment_boundary(request_queue_t *q, unsigned long mask)
760 if (mask < PAGE_CACHE_SIZE - 1) {
761 mask = PAGE_CACHE_SIZE - 1;
762 printk("%s: set to minimum %lx\n", __FUNCTION__, mask);
765 q->seg_boundary_mask = mask;
768 EXPORT_SYMBOL(blk_queue_segment_boundary);
771 * blk_queue_dma_alignment - set dma length and memory alignment
772 * @q: the request queue for the device
773 * @mask: alignment mask
775 * description:
776 * set required memory and length aligment for direct dma transactions.
777 * this is used when buiding direct io requests for the queue.
780 void blk_queue_dma_alignment(request_queue_t *q, int mask)
782 q->dma_alignment = mask;
785 EXPORT_SYMBOL(blk_queue_dma_alignment);
788 * blk_queue_find_tag - find a request by its tag and queue
789 * @q: The request queue for the device
790 * @tag: The tag of the request
792 * Notes:
793 * Should be used when a device returns a tag and you want to match
794 * it with a request.
796 * no locks need be held.
798 struct request *blk_queue_find_tag(request_queue_t *q, int tag)
800 return blk_map_queue_find_tag(q->queue_tags, tag);
803 EXPORT_SYMBOL(blk_queue_find_tag);
806 * __blk_free_tags - release a given set of tag maintenance info
807 * @bqt: the tag map to free
809 * Tries to free the specified @bqt@. Returns true if it was
810 * actually freed and false if there are still references using it
812 static int __blk_free_tags(struct blk_queue_tag *bqt)
814 int retval;
816 retval = atomic_dec_and_test(&bqt->refcnt);
817 if (retval) {
818 BUG_ON(bqt->busy);
819 BUG_ON(!list_empty(&bqt->busy_list));
821 kfree(bqt->tag_index);
822 bqt->tag_index = NULL;
824 kfree(bqt->tag_map);
825 bqt->tag_map = NULL;
827 kfree(bqt);
831 return retval;
835 * __blk_queue_free_tags - release tag maintenance info
836 * @q: the request queue for the device
838 * Notes:
839 * blk_cleanup_queue() will take care of calling this function, if tagging
840 * has been used. So there's no need to call this directly.
842 static void __blk_queue_free_tags(request_queue_t *q)
844 struct blk_queue_tag *bqt = q->queue_tags;
846 if (!bqt)
847 return;
849 __blk_free_tags(bqt);
851 q->queue_tags = NULL;
852 q->queue_flags &= ~(1 << QUEUE_FLAG_QUEUED);
857 * blk_free_tags - release a given set of tag maintenance info
858 * @bqt: the tag map to free
860 * For externally managed @bqt@ frees the map. Callers of this
861 * function must guarantee to have released all the queues that
862 * might have been using this tag map.
864 void blk_free_tags(struct blk_queue_tag *bqt)
866 if (unlikely(!__blk_free_tags(bqt)))
867 BUG();
869 EXPORT_SYMBOL(blk_free_tags);
872 * blk_queue_free_tags - release tag maintenance info
873 * @q: the request queue for the device
875 * Notes:
876 * This is used to disabled tagged queuing to a device, yet leave
877 * queue in function.
879 void blk_queue_free_tags(request_queue_t *q)
881 clear_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
884 EXPORT_SYMBOL(blk_queue_free_tags);
886 static int
887 init_tag_map(request_queue_t *q, struct blk_queue_tag *tags, int depth)
889 struct request **tag_index;
890 unsigned long *tag_map;
891 int nr_ulongs;
893 if (q && depth > q->nr_requests * 2) {
894 depth = q->nr_requests * 2;
895 printk(KERN_ERR "%s: adjusted depth to %d\n",
896 __FUNCTION__, depth);
899 tag_index = kzalloc(depth * sizeof(struct request *), GFP_ATOMIC);
900 if (!tag_index)
901 goto fail;
903 nr_ulongs = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG;
904 tag_map = kzalloc(nr_ulongs * sizeof(unsigned long), GFP_ATOMIC);
905 if (!tag_map)
906 goto fail;
908 tags->real_max_depth = depth;
909 tags->max_depth = depth;
910 tags->tag_index = tag_index;
911 tags->tag_map = tag_map;
913 return 0;
914 fail:
915 kfree(tag_index);
916 return -ENOMEM;
919 static struct blk_queue_tag *__blk_queue_init_tags(struct request_queue *q,
920 int depth)
922 struct blk_queue_tag *tags;
924 tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC);
925 if (!tags)
926 goto fail;
928 if (init_tag_map(q, tags, depth))
929 goto fail;
931 INIT_LIST_HEAD(&tags->busy_list);
932 tags->busy = 0;
933 atomic_set(&tags->refcnt, 1);
934 return tags;
935 fail:
936 kfree(tags);
937 return NULL;
941 * blk_init_tags - initialize the tag info for an external tag map
942 * @depth: the maximum queue depth supported
943 * @tags: the tag to use
945 struct blk_queue_tag *blk_init_tags(int depth)
947 return __blk_queue_init_tags(NULL, depth);
949 EXPORT_SYMBOL(blk_init_tags);
952 * blk_queue_init_tags - initialize the queue tag info
953 * @q: the request queue for the device
954 * @depth: the maximum queue depth supported
955 * @tags: the tag to use
957 int blk_queue_init_tags(request_queue_t *q, int depth,
958 struct blk_queue_tag *tags)
960 int rc;
962 BUG_ON(tags && q->queue_tags && tags != q->queue_tags);
964 if (!tags && !q->queue_tags) {
965 tags = __blk_queue_init_tags(q, depth);
967 if (!tags)
968 goto fail;
969 } else if (q->queue_tags) {
970 if ((rc = blk_queue_resize_tags(q, depth)))
971 return rc;
972 set_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
973 return 0;
974 } else
975 atomic_inc(&tags->refcnt);
978 * assign it, all done
980 q->queue_tags = tags;
981 q->queue_flags |= (1 << QUEUE_FLAG_QUEUED);
982 return 0;
983 fail:
984 kfree(tags);
985 return -ENOMEM;
988 EXPORT_SYMBOL(blk_queue_init_tags);
991 * blk_queue_resize_tags - change the queueing depth
992 * @q: the request queue for the device
993 * @new_depth: the new max command queueing depth
995 * Notes:
996 * Must be called with the queue lock held.
998 int blk_queue_resize_tags(request_queue_t *q, int new_depth)
1000 struct blk_queue_tag *bqt = q->queue_tags;
1001 struct request **tag_index;
1002 unsigned long *tag_map;
1003 int max_depth, nr_ulongs;
1005 if (!bqt)
1006 return -ENXIO;
1009 * if we already have large enough real_max_depth. just
1010 * adjust max_depth. *NOTE* as requests with tag value
1011 * between new_depth and real_max_depth can be in-flight, tag
1012 * map can not be shrunk blindly here.
1014 if (new_depth <= bqt->real_max_depth) {
1015 bqt->max_depth = new_depth;
1016 return 0;
1020 * Currently cannot replace a shared tag map with a new
1021 * one, so error out if this is the case
1023 if (atomic_read(&bqt->refcnt) != 1)
1024 return -EBUSY;
1027 * save the old state info, so we can copy it back
1029 tag_index = bqt->tag_index;
1030 tag_map = bqt->tag_map;
1031 max_depth = bqt->real_max_depth;
1033 if (init_tag_map(q, bqt, new_depth))
1034 return -ENOMEM;
1036 memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
1037 nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG;
1038 memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long));
1040 kfree(tag_index);
1041 kfree(tag_map);
1042 return 0;
1045 EXPORT_SYMBOL(blk_queue_resize_tags);
1048 * blk_queue_end_tag - end tag operations for a request
1049 * @q: the request queue for the device
1050 * @rq: the request that has completed
1052 * Description:
1053 * Typically called when end_that_request_first() returns 0, meaning
1054 * all transfers have been done for a request. It's important to call
1055 * this function before end_that_request_last(), as that will put the
1056 * request back on the free list thus corrupting the internal tag list.
1058 * Notes:
1059 * queue lock must be held.
1061 void blk_queue_end_tag(request_queue_t *q, struct request *rq)
1063 struct blk_queue_tag *bqt = q->queue_tags;
1064 int tag = rq->tag;
1066 BUG_ON(tag == -1);
1068 if (unlikely(tag >= bqt->real_max_depth))
1070 * This can happen after tag depth has been reduced.
1071 * FIXME: how about a warning or info message here?
1073 return;
1075 if (unlikely(!__test_and_clear_bit(tag, bqt->tag_map))) {
1076 printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n",
1077 __FUNCTION__, tag);
1078 return;
1081 list_del_init(&rq->queuelist);
1082 rq->cmd_flags &= ~REQ_QUEUED;
1083 rq->tag = -1;
1085 if (unlikely(bqt->tag_index[tag] == NULL))
1086 printk(KERN_ERR "%s: tag %d is missing\n",
1087 __FUNCTION__, tag);
1089 bqt->tag_index[tag] = NULL;
1090 bqt->busy--;
1093 EXPORT_SYMBOL(blk_queue_end_tag);
1096 * blk_queue_start_tag - find a free tag and assign it
1097 * @q: the request queue for the device
1098 * @rq: the block request that needs tagging
1100 * Description:
1101 * This can either be used as a stand-alone helper, or possibly be
1102 * assigned as the queue &prep_rq_fn (in which case &struct request
1103 * automagically gets a tag assigned). Note that this function
1104 * assumes that any type of request can be queued! if this is not
1105 * true for your device, you must check the request type before
1106 * calling this function. The request will also be removed from
1107 * the request queue, so it's the drivers responsibility to readd
1108 * it if it should need to be restarted for some reason.
1110 * Notes:
1111 * queue lock must be held.
1113 int blk_queue_start_tag(request_queue_t *q, struct request *rq)
1115 struct blk_queue_tag *bqt = q->queue_tags;
1116 int tag;
1118 if (unlikely((rq->cmd_flags & REQ_QUEUED))) {
1119 printk(KERN_ERR
1120 "%s: request %p for device [%s] already tagged %d",
1121 __FUNCTION__, rq,
1122 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag);
1123 BUG();
1127 * Protect against shared tag maps, as we may not have exclusive
1128 * access to the tag map.
1130 do {
1131 tag = find_first_zero_bit(bqt->tag_map, bqt->max_depth);
1132 if (tag >= bqt->max_depth)
1133 return 1;
1135 } while (test_and_set_bit(tag, bqt->tag_map));
1137 rq->cmd_flags |= REQ_QUEUED;
1138 rq->tag = tag;
1139 bqt->tag_index[tag] = rq;
1140 blkdev_dequeue_request(rq);
1141 list_add(&rq->queuelist, &bqt->busy_list);
1142 bqt->busy++;
1143 return 0;
1146 EXPORT_SYMBOL(blk_queue_start_tag);
1149 * blk_queue_invalidate_tags - invalidate all pending tags
1150 * @q: the request queue for the device
1152 * Description:
1153 * Hardware conditions may dictate a need to stop all pending requests.
1154 * In this case, we will safely clear the block side of the tag queue and
1155 * readd all requests to the request queue in the right order.
1157 * Notes:
1158 * queue lock must be held.
1160 void blk_queue_invalidate_tags(request_queue_t *q)
1162 struct blk_queue_tag *bqt = q->queue_tags;
1163 struct list_head *tmp, *n;
1164 struct request *rq;
1166 list_for_each_safe(tmp, n, &bqt->busy_list) {
1167 rq = list_entry_rq(tmp);
1169 if (rq->tag == -1) {
1170 printk(KERN_ERR
1171 "%s: bad tag found on list\n", __FUNCTION__);
1172 list_del_init(&rq->queuelist);
1173 rq->cmd_flags &= ~REQ_QUEUED;
1174 } else
1175 blk_queue_end_tag(q, rq);
1177 rq->cmd_flags &= ~REQ_STARTED;
1178 __elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 0);
1182 EXPORT_SYMBOL(blk_queue_invalidate_tags);
1184 void blk_dump_rq_flags(struct request *rq, char *msg)
1186 int bit;
1188 printk("%s: dev %s: type=%x, flags=%x\n", msg,
1189 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->cmd_type,
1190 rq->cmd_flags);
1192 printk("\nsector %llu, nr/cnr %lu/%u\n", (unsigned long long)rq->sector,
1193 rq->nr_sectors,
1194 rq->current_nr_sectors);
1195 printk("bio %p, biotail %p, buffer %p, data %p, len %u\n", rq->bio, rq->biotail, rq->buffer, rq->data, rq->data_len);
1197 if (blk_pc_request(rq)) {
1198 printk("cdb: ");
1199 for (bit = 0; bit < sizeof(rq->cmd); bit++)
1200 printk("%02x ", rq->cmd[bit]);
1201 printk("\n");
1205 EXPORT_SYMBOL(blk_dump_rq_flags);
1207 void blk_recount_segments(request_queue_t *q, struct bio *bio)
1209 struct bio_vec *bv, *bvprv = NULL;
1210 int i, nr_phys_segs, nr_hw_segs, seg_size, hw_seg_size, cluster;
1211 int high, highprv = 1;
1213 if (unlikely(!bio->bi_io_vec))
1214 return;
1216 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1217 hw_seg_size = seg_size = nr_phys_segs = nr_hw_segs = 0;
1218 bio_for_each_segment(bv, bio, i) {
1220 * the trick here is making sure that a high page is never
1221 * considered part of another segment, since that might
1222 * change with the bounce page.
1224 high = page_to_pfn(bv->bv_page) > q->bounce_pfn;
1225 if (high || highprv)
1226 goto new_hw_segment;
1227 if (cluster) {
1228 if (seg_size + bv->bv_len > q->max_segment_size)
1229 goto new_segment;
1230 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
1231 goto new_segment;
1232 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
1233 goto new_segment;
1234 if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1235 goto new_hw_segment;
1237 seg_size += bv->bv_len;
1238 hw_seg_size += bv->bv_len;
1239 bvprv = bv;
1240 continue;
1242 new_segment:
1243 if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
1244 !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len)) {
1245 hw_seg_size += bv->bv_len;
1246 } else {
1247 new_hw_segment:
1248 if (hw_seg_size > bio->bi_hw_front_size)
1249 bio->bi_hw_front_size = hw_seg_size;
1250 hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
1251 nr_hw_segs++;
1254 nr_phys_segs++;
1255 bvprv = bv;
1256 seg_size = bv->bv_len;
1257 highprv = high;
1259 if (hw_seg_size > bio->bi_hw_back_size)
1260 bio->bi_hw_back_size = hw_seg_size;
1261 if (nr_hw_segs == 1 && hw_seg_size > bio->bi_hw_front_size)
1262 bio->bi_hw_front_size = hw_seg_size;
1263 bio->bi_phys_segments = nr_phys_segs;
1264 bio->bi_hw_segments = nr_hw_segs;
1265 bio->bi_flags |= (1 << BIO_SEG_VALID);
1267 EXPORT_SYMBOL(blk_recount_segments);
1269 static int blk_phys_contig_segment(request_queue_t *q, struct bio *bio,
1270 struct bio *nxt)
1272 if (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
1273 return 0;
1275 if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
1276 return 0;
1277 if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1278 return 0;
1281 * bio and nxt are contigous in memory, check if the queue allows
1282 * these two to be merged into one
1284 if (BIO_SEG_BOUNDARY(q, bio, nxt))
1285 return 1;
1287 return 0;
1290 static int blk_hw_contig_segment(request_queue_t *q, struct bio *bio,
1291 struct bio *nxt)
1293 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1294 blk_recount_segments(q, bio);
1295 if (unlikely(!bio_flagged(nxt, BIO_SEG_VALID)))
1296 blk_recount_segments(q, nxt);
1297 if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
1298 BIOVEC_VIRT_OVERSIZE(bio->bi_hw_front_size + bio->bi_hw_back_size))
1299 return 0;
1300 if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1301 return 0;
1303 return 1;
1307 * map a request to scatterlist, return number of sg entries setup. Caller
1308 * must make sure sg can hold rq->nr_phys_segments entries
1310 int blk_rq_map_sg(request_queue_t *q, struct request *rq, struct scatterlist *sg)
1312 struct bio_vec *bvec, *bvprv;
1313 struct bio *bio;
1314 int nsegs, i, cluster;
1316 nsegs = 0;
1317 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1320 * for each bio in rq
1322 bvprv = NULL;
1323 rq_for_each_bio(bio, rq) {
1325 * for each segment in bio
1327 bio_for_each_segment(bvec, bio, i) {
1328 int nbytes = bvec->bv_len;
1330 if (bvprv && cluster) {
1331 if (sg[nsegs - 1].length + nbytes > q->max_segment_size)
1332 goto new_segment;
1334 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
1335 goto new_segment;
1336 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
1337 goto new_segment;
1339 sg[nsegs - 1].length += nbytes;
1340 } else {
1341 new_segment:
1342 memset(&sg[nsegs],0,sizeof(struct scatterlist));
1343 sg[nsegs].page = bvec->bv_page;
1344 sg[nsegs].length = nbytes;
1345 sg[nsegs].offset = bvec->bv_offset;
1347 nsegs++;
1349 bvprv = bvec;
1350 } /* segments in bio */
1351 } /* bios in rq */
1353 return nsegs;
1356 EXPORT_SYMBOL(blk_rq_map_sg);
1359 * the standard queue merge functions, can be overridden with device
1360 * specific ones if so desired
1363 static inline int ll_new_mergeable(request_queue_t *q,
1364 struct request *req,
1365 struct bio *bio)
1367 int nr_phys_segs = bio_phys_segments(q, bio);
1369 if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1370 req->cmd_flags |= REQ_NOMERGE;
1371 if (req == q->last_merge)
1372 q->last_merge = NULL;
1373 return 0;
1377 * A hw segment is just getting larger, bump just the phys
1378 * counter.
1380 req->nr_phys_segments += nr_phys_segs;
1381 return 1;
1384 static inline int ll_new_hw_segment(request_queue_t *q,
1385 struct request *req,
1386 struct bio *bio)
1388 int nr_hw_segs = bio_hw_segments(q, bio);
1389 int nr_phys_segs = bio_phys_segments(q, bio);
1391 if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
1392 || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1393 req->cmd_flags |= REQ_NOMERGE;
1394 if (req == q->last_merge)
1395 q->last_merge = NULL;
1396 return 0;
1400 * This will form the start of a new hw segment. Bump both
1401 * counters.
1403 req->nr_hw_segments += nr_hw_segs;
1404 req->nr_phys_segments += nr_phys_segs;
1405 return 1;
1408 int ll_back_merge_fn(request_queue_t *q, struct request *req, struct bio *bio)
1410 unsigned short max_sectors;
1411 int len;
1413 if (unlikely(blk_pc_request(req)))
1414 max_sectors = q->max_hw_sectors;
1415 else
1416 max_sectors = q->max_sectors;
1418 if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1419 req->cmd_flags |= REQ_NOMERGE;
1420 if (req == q->last_merge)
1421 q->last_merge = NULL;
1422 return 0;
1424 if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
1425 blk_recount_segments(q, req->biotail);
1426 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1427 blk_recount_segments(q, bio);
1428 len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
1429 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio)) &&
1430 !BIOVEC_VIRT_OVERSIZE(len)) {
1431 int mergeable = ll_new_mergeable(q, req, bio);
1433 if (mergeable) {
1434 if (req->nr_hw_segments == 1)
1435 req->bio->bi_hw_front_size = len;
1436 if (bio->bi_hw_segments == 1)
1437 bio->bi_hw_back_size = len;
1439 return mergeable;
1442 return ll_new_hw_segment(q, req, bio);
1444 EXPORT_SYMBOL(ll_back_merge_fn);
1446 static int ll_front_merge_fn(request_queue_t *q, struct request *req,
1447 struct bio *bio)
1449 unsigned short max_sectors;
1450 int len;
1452 if (unlikely(blk_pc_request(req)))
1453 max_sectors = q->max_hw_sectors;
1454 else
1455 max_sectors = q->max_sectors;
1458 if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1459 req->cmd_flags |= REQ_NOMERGE;
1460 if (req == q->last_merge)
1461 q->last_merge = NULL;
1462 return 0;
1464 len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
1465 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1466 blk_recount_segments(q, bio);
1467 if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
1468 blk_recount_segments(q, req->bio);
1469 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
1470 !BIOVEC_VIRT_OVERSIZE(len)) {
1471 int mergeable = ll_new_mergeable(q, req, bio);
1473 if (mergeable) {
1474 if (bio->bi_hw_segments == 1)
1475 bio->bi_hw_front_size = len;
1476 if (req->nr_hw_segments == 1)
1477 req->biotail->bi_hw_back_size = len;
1479 return mergeable;
1482 return ll_new_hw_segment(q, req, bio);
1485 static int ll_merge_requests_fn(request_queue_t *q, struct request *req,
1486 struct request *next)
1488 int total_phys_segments;
1489 int total_hw_segments;
1492 * First check if the either of the requests are re-queued
1493 * requests. Can't merge them if they are.
1495 if (req->special || next->special)
1496 return 0;
1499 * Will it become too large?
1501 if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
1502 return 0;
1504 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
1505 if (blk_phys_contig_segment(q, req->biotail, next->bio))
1506 total_phys_segments--;
1508 if (total_phys_segments > q->max_phys_segments)
1509 return 0;
1511 total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1512 if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
1513 int len = req->biotail->bi_hw_back_size + next->bio->bi_hw_front_size;
1515 * propagate the combined length to the end of the requests
1517 if (req->nr_hw_segments == 1)
1518 req->bio->bi_hw_front_size = len;
1519 if (next->nr_hw_segments == 1)
1520 next->biotail->bi_hw_back_size = len;
1521 total_hw_segments--;
1524 if (total_hw_segments > q->max_hw_segments)
1525 return 0;
1527 /* Merge is OK... */
1528 req->nr_phys_segments = total_phys_segments;
1529 req->nr_hw_segments = total_hw_segments;
1530 return 1;
1534 * "plug" the device if there are no outstanding requests: this will
1535 * force the transfer to start only after we have put all the requests
1536 * on the list.
1538 * This is called with interrupts off and no requests on the queue and
1539 * with the queue lock held.
1541 void blk_plug_device(request_queue_t *q)
1543 WARN_ON(!irqs_disabled());
1546 * don't plug a stopped queue, it must be paired with blk_start_queue()
1547 * which will restart the queueing
1549 if (blk_queue_stopped(q))
1550 return;
1552 if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags)) {
1553 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
1554 blk_add_trace_generic(q, NULL, 0, BLK_TA_PLUG);
1558 EXPORT_SYMBOL(blk_plug_device);
1561 * remove the queue from the plugged list, if present. called with
1562 * queue lock held and interrupts disabled.
1564 int blk_remove_plug(request_queue_t *q)
1566 WARN_ON(!irqs_disabled());
1568 if (!test_and_clear_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1569 return 0;
1571 del_timer(&q->unplug_timer);
1572 return 1;
1575 EXPORT_SYMBOL(blk_remove_plug);
1578 * remove the plug and let it rip..
1580 void __generic_unplug_device(request_queue_t *q)
1582 if (unlikely(blk_queue_stopped(q)))
1583 return;
1585 if (!blk_remove_plug(q))
1586 return;
1588 q->request_fn(q);
1590 EXPORT_SYMBOL(__generic_unplug_device);
1593 * generic_unplug_device - fire a request queue
1594 * @q: The &request_queue_t in question
1596 * Description:
1597 * Linux uses plugging to build bigger requests queues before letting
1598 * the device have at them. If a queue is plugged, the I/O scheduler
1599 * is still adding and merging requests on the queue. Once the queue
1600 * gets unplugged, the request_fn defined for the queue is invoked and
1601 * transfers started.
1603 void generic_unplug_device(request_queue_t *q)
1605 spin_lock_irq(q->queue_lock);
1606 __generic_unplug_device(q);
1607 spin_unlock_irq(q->queue_lock);
1609 EXPORT_SYMBOL(generic_unplug_device);
1611 static void blk_backing_dev_unplug(struct backing_dev_info *bdi,
1612 struct page *page)
1614 request_queue_t *q = bdi->unplug_io_data;
1617 * devices don't necessarily have an ->unplug_fn defined
1619 if (q->unplug_fn) {
1620 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1621 q->rq.count[READ] + q->rq.count[WRITE]);
1623 q->unplug_fn(q);
1627 static void blk_unplug_work(struct work_struct *work)
1629 request_queue_t *q = container_of(work, request_queue_t, unplug_work);
1631 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1632 q->rq.count[READ] + q->rq.count[WRITE]);
1634 q->unplug_fn(q);
1637 static void blk_unplug_timeout(unsigned long data)
1639 request_queue_t *q = (request_queue_t *)data;
1641 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_TIMER, NULL,
1642 q->rq.count[READ] + q->rq.count[WRITE]);
1644 kblockd_schedule_work(&q->unplug_work);
1648 * blk_start_queue - restart a previously stopped queue
1649 * @q: The &request_queue_t in question
1651 * Description:
1652 * blk_start_queue() will clear the stop flag on the queue, and call
1653 * the request_fn for the queue if it was in a stopped state when
1654 * entered. Also see blk_stop_queue(). Queue lock must be held.
1656 void blk_start_queue(request_queue_t *q)
1658 WARN_ON(!irqs_disabled());
1660 clear_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1663 * one level of recursion is ok and is much faster than kicking
1664 * the unplug handling
1666 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1667 q->request_fn(q);
1668 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1669 } else {
1670 blk_plug_device(q);
1671 kblockd_schedule_work(&q->unplug_work);
1675 EXPORT_SYMBOL(blk_start_queue);
1678 * blk_stop_queue - stop a queue
1679 * @q: The &request_queue_t in question
1681 * Description:
1682 * The Linux block layer assumes that a block driver will consume all
1683 * entries on the request queue when the request_fn strategy is called.
1684 * Often this will not happen, because of hardware limitations (queue
1685 * depth settings). If a device driver gets a 'queue full' response,
1686 * or if it simply chooses not to queue more I/O at one point, it can
1687 * call this function to prevent the request_fn from being called until
1688 * the driver has signalled it's ready to go again. This happens by calling
1689 * blk_start_queue() to restart queue operations. Queue lock must be held.
1691 void blk_stop_queue(request_queue_t *q)
1693 blk_remove_plug(q);
1694 set_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1696 EXPORT_SYMBOL(blk_stop_queue);
1699 * blk_sync_queue - cancel any pending callbacks on a queue
1700 * @q: the queue
1702 * Description:
1703 * The block layer may perform asynchronous callback activity
1704 * on a queue, such as calling the unplug function after a timeout.
1705 * A block device may call blk_sync_queue to ensure that any
1706 * such activity is cancelled, thus allowing it to release resources
1707 * the the callbacks might use. The caller must already have made sure
1708 * that its ->make_request_fn will not re-add plugging prior to calling
1709 * this function.
1712 void blk_sync_queue(struct request_queue *q)
1714 del_timer_sync(&q->unplug_timer);
1715 kblockd_flush();
1717 EXPORT_SYMBOL(blk_sync_queue);
1720 * blk_run_queue - run a single device queue
1721 * @q: The queue to run
1723 void blk_run_queue(struct request_queue *q)
1725 unsigned long flags;
1727 spin_lock_irqsave(q->queue_lock, flags);
1728 blk_remove_plug(q);
1731 * Only recurse once to avoid overrunning the stack, let the unplug
1732 * handling reinvoke the handler shortly if we already got there.
1734 if (!elv_queue_empty(q)) {
1735 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1736 q->request_fn(q);
1737 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1738 } else {
1739 blk_plug_device(q);
1740 kblockd_schedule_work(&q->unplug_work);
1744 spin_unlock_irqrestore(q->queue_lock, flags);
1746 EXPORT_SYMBOL(blk_run_queue);
1749 * blk_cleanup_queue: - release a &request_queue_t when it is no longer needed
1750 * @kobj: the kobj belonging of the request queue to be released
1752 * Description:
1753 * blk_cleanup_queue is the pair to blk_init_queue() or
1754 * blk_queue_make_request(). It should be called when a request queue is
1755 * being released; typically when a block device is being de-registered.
1756 * Currently, its primary task it to free all the &struct request
1757 * structures that were allocated to the queue and the queue itself.
1759 * Caveat:
1760 * Hopefully the low level driver will have finished any
1761 * outstanding requests first...
1763 static void blk_release_queue(struct kobject *kobj)
1765 request_queue_t *q = container_of(kobj, struct request_queue, kobj);
1766 struct request_list *rl = &q->rq;
1768 blk_sync_queue(q);
1770 if (rl->rq_pool)
1771 mempool_destroy(rl->rq_pool);
1773 if (q->queue_tags)
1774 __blk_queue_free_tags(q);
1776 blk_trace_shutdown(q);
1778 kmem_cache_free(requestq_cachep, q);
1781 void blk_put_queue(request_queue_t *q)
1783 kobject_put(&q->kobj);
1785 EXPORT_SYMBOL(blk_put_queue);
1787 void blk_cleanup_queue(request_queue_t * q)
1789 mutex_lock(&q->sysfs_lock);
1790 set_bit(QUEUE_FLAG_DEAD, &q->queue_flags);
1791 mutex_unlock(&q->sysfs_lock);
1793 if (q->elevator)
1794 elevator_exit(q->elevator);
1796 blk_put_queue(q);
1799 EXPORT_SYMBOL(blk_cleanup_queue);
1801 static int blk_init_free_list(request_queue_t *q)
1803 struct request_list *rl = &q->rq;
1805 rl->count[READ] = rl->count[WRITE] = 0;
1806 rl->starved[READ] = rl->starved[WRITE] = 0;
1807 rl->elvpriv = 0;
1808 init_waitqueue_head(&rl->wait[READ]);
1809 init_waitqueue_head(&rl->wait[WRITE]);
1811 rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
1812 mempool_free_slab, request_cachep, q->node);
1814 if (!rl->rq_pool)
1815 return -ENOMEM;
1817 return 0;
1820 request_queue_t *blk_alloc_queue(gfp_t gfp_mask)
1822 return blk_alloc_queue_node(gfp_mask, -1);
1824 EXPORT_SYMBOL(blk_alloc_queue);
1826 static struct kobj_type queue_ktype;
1828 request_queue_t *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
1830 request_queue_t *q;
1832 q = kmem_cache_alloc_node(requestq_cachep, gfp_mask, node_id);
1833 if (!q)
1834 return NULL;
1836 memset(q, 0, sizeof(*q));
1837 init_timer(&q->unplug_timer);
1839 snprintf(q->kobj.name, KOBJ_NAME_LEN, "%s", "queue");
1840 q->kobj.ktype = &queue_ktype;
1841 kobject_init(&q->kobj);
1843 q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
1844 q->backing_dev_info.unplug_io_data = q;
1846 mutex_init(&q->sysfs_lock);
1848 return q;
1850 EXPORT_SYMBOL(blk_alloc_queue_node);
1853 * blk_init_queue - prepare a request queue for use with a block device
1854 * @rfn: The function to be called to process requests that have been
1855 * placed on the queue.
1856 * @lock: Request queue spin lock
1858 * Description:
1859 * If a block device wishes to use the standard request handling procedures,
1860 * which sorts requests and coalesces adjacent requests, then it must
1861 * call blk_init_queue(). The function @rfn will be called when there
1862 * are requests on the queue that need to be processed. If the device
1863 * supports plugging, then @rfn may not be called immediately when requests
1864 * are available on the queue, but may be called at some time later instead.
1865 * Plugged queues are generally unplugged when a buffer belonging to one
1866 * of the requests on the queue is needed, or due to memory pressure.
1868 * @rfn is not required, or even expected, to remove all requests off the
1869 * queue, but only as many as it can handle at a time. If it does leave
1870 * requests on the queue, it is responsible for arranging that the requests
1871 * get dealt with eventually.
1873 * The queue spin lock must be held while manipulating the requests on the
1874 * request queue; this lock will be taken also from interrupt context, so irq
1875 * disabling is needed for it.
1877 * Function returns a pointer to the initialized request queue, or NULL if
1878 * it didn't succeed.
1880 * Note:
1881 * blk_init_queue() must be paired with a blk_cleanup_queue() call
1882 * when the block device is deactivated (such as at module unload).
1885 request_queue_t *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
1887 return blk_init_queue_node(rfn, lock, -1);
1889 EXPORT_SYMBOL(blk_init_queue);
1891 request_queue_t *
1892 blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
1894 request_queue_t *q = blk_alloc_queue_node(GFP_KERNEL, node_id);
1896 if (!q)
1897 return NULL;
1899 q->node = node_id;
1900 if (blk_init_free_list(q)) {
1901 kmem_cache_free(requestq_cachep, q);
1902 return NULL;
1906 * if caller didn't supply a lock, they get per-queue locking with
1907 * our embedded lock
1909 if (!lock) {
1910 spin_lock_init(&q->__queue_lock);
1911 lock = &q->__queue_lock;
1914 q->request_fn = rfn;
1915 q->prep_rq_fn = NULL;
1916 q->unplug_fn = generic_unplug_device;
1917 q->queue_flags = (1 << QUEUE_FLAG_CLUSTER);
1918 q->queue_lock = lock;
1920 blk_queue_segment_boundary(q, 0xffffffff);
1922 blk_queue_make_request(q, __make_request);
1923 blk_queue_max_segment_size(q, MAX_SEGMENT_SIZE);
1925 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
1926 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
1929 * all done
1931 if (!elevator_init(q, NULL)) {
1932 blk_queue_congestion_threshold(q);
1933 return q;
1936 blk_put_queue(q);
1937 return NULL;
1939 EXPORT_SYMBOL(blk_init_queue_node);
1941 int blk_get_queue(request_queue_t *q)
1943 if (likely(!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
1944 kobject_get(&q->kobj);
1945 return 0;
1948 return 1;
1951 EXPORT_SYMBOL(blk_get_queue);
1953 static inline void blk_free_request(request_queue_t *q, struct request *rq)
1955 if (rq->cmd_flags & REQ_ELVPRIV)
1956 elv_put_request(q, rq);
1957 mempool_free(rq, q->rq.rq_pool);
1960 static struct request *
1961 blk_alloc_request(request_queue_t *q, int rw, int priv, gfp_t gfp_mask)
1963 struct request *rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
1965 if (!rq)
1966 return NULL;
1969 * first three bits are identical in rq->cmd_flags and bio->bi_rw,
1970 * see bio.h and blkdev.h
1972 rq->cmd_flags = rw | REQ_ALLOCED;
1974 if (priv) {
1975 if (unlikely(elv_set_request(q, rq, gfp_mask))) {
1976 mempool_free(rq, q->rq.rq_pool);
1977 return NULL;
1979 rq->cmd_flags |= REQ_ELVPRIV;
1982 return rq;
1986 * ioc_batching returns true if the ioc is a valid batching request and
1987 * should be given priority access to a request.
1989 static inline int ioc_batching(request_queue_t *q, struct io_context *ioc)
1991 if (!ioc)
1992 return 0;
1995 * Make sure the process is able to allocate at least 1 request
1996 * even if the batch times out, otherwise we could theoretically
1997 * lose wakeups.
1999 return ioc->nr_batch_requests == q->nr_batching ||
2000 (ioc->nr_batch_requests > 0
2001 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
2005 * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
2006 * will cause the process to be a "batcher" on all queues in the system. This
2007 * is the behaviour we want though - once it gets a wakeup it should be given
2008 * a nice run.
2010 static void ioc_set_batching(request_queue_t *q, struct io_context *ioc)
2012 if (!ioc || ioc_batching(q, ioc))
2013 return;
2015 ioc->nr_batch_requests = q->nr_batching;
2016 ioc->last_waited = jiffies;
2019 static void __freed_request(request_queue_t *q, int rw)
2021 struct request_list *rl = &q->rq;
2023 if (rl->count[rw] < queue_congestion_off_threshold(q))
2024 blk_clear_queue_congested(q, rw);
2026 if (rl->count[rw] + 1 <= q->nr_requests) {
2027 if (waitqueue_active(&rl->wait[rw]))
2028 wake_up(&rl->wait[rw]);
2030 blk_clear_queue_full(q, rw);
2035 * A request has just been released. Account for it, update the full and
2036 * congestion status, wake up any waiters. Called under q->queue_lock.
2038 static void freed_request(request_queue_t *q, int rw, int priv)
2040 struct request_list *rl = &q->rq;
2042 rl->count[rw]--;
2043 if (priv)
2044 rl->elvpriv--;
2046 __freed_request(q, rw);
2048 if (unlikely(rl->starved[rw ^ 1]))
2049 __freed_request(q, rw ^ 1);
2052 #define blkdev_free_rq(list) list_entry((list)->next, struct request, queuelist)
2054 * Get a free request, queue_lock must be held.
2055 * Returns NULL on failure, with queue_lock held.
2056 * Returns !NULL on success, with queue_lock *not held*.
2058 static struct request *get_request(request_queue_t *q, int rw_flags,
2059 struct bio *bio, gfp_t gfp_mask)
2061 struct request *rq = NULL;
2062 struct request_list *rl = &q->rq;
2063 struct io_context *ioc = NULL;
2064 const int rw = rw_flags & 0x01;
2065 int may_queue, priv;
2067 may_queue = elv_may_queue(q, rw_flags);
2068 if (may_queue == ELV_MQUEUE_NO)
2069 goto rq_starved;
2071 if (rl->count[rw]+1 >= queue_congestion_on_threshold(q)) {
2072 if (rl->count[rw]+1 >= q->nr_requests) {
2073 ioc = current_io_context(GFP_ATOMIC, q->node);
2075 * The queue will fill after this allocation, so set
2076 * it as full, and mark this process as "batching".
2077 * This process will be allowed to complete a batch of
2078 * requests, others will be blocked.
2080 if (!blk_queue_full(q, rw)) {
2081 ioc_set_batching(q, ioc);
2082 blk_set_queue_full(q, rw);
2083 } else {
2084 if (may_queue != ELV_MQUEUE_MUST
2085 && !ioc_batching(q, ioc)) {
2087 * The queue is full and the allocating
2088 * process is not a "batcher", and not
2089 * exempted by the IO scheduler
2091 goto out;
2095 blk_set_queue_congested(q, rw);
2099 * Only allow batching queuers to allocate up to 50% over the defined
2100 * limit of requests, otherwise we could have thousands of requests
2101 * allocated with any setting of ->nr_requests
2103 if (rl->count[rw] >= (3 * q->nr_requests / 2))
2104 goto out;
2106 rl->count[rw]++;
2107 rl->starved[rw] = 0;
2109 priv = !test_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
2110 if (priv)
2111 rl->elvpriv++;
2113 spin_unlock_irq(q->queue_lock);
2115 rq = blk_alloc_request(q, rw_flags, priv, gfp_mask);
2116 if (unlikely(!rq)) {
2118 * Allocation failed presumably due to memory. Undo anything
2119 * we might have messed up.
2121 * Allocating task should really be put onto the front of the
2122 * wait queue, but this is pretty rare.
2124 spin_lock_irq(q->queue_lock);
2125 freed_request(q, rw, priv);
2128 * in the very unlikely event that allocation failed and no
2129 * requests for this direction was pending, mark us starved
2130 * so that freeing of a request in the other direction will
2131 * notice us. another possible fix would be to split the
2132 * rq mempool into READ and WRITE
2134 rq_starved:
2135 if (unlikely(rl->count[rw] == 0))
2136 rl->starved[rw] = 1;
2138 goto out;
2142 * ioc may be NULL here, and ioc_batching will be false. That's
2143 * OK, if the queue is under the request limit then requests need
2144 * not count toward the nr_batch_requests limit. There will always
2145 * be some limit enforced by BLK_BATCH_TIME.
2147 if (ioc_batching(q, ioc))
2148 ioc->nr_batch_requests--;
2150 rq_init(q, rq);
2152 blk_add_trace_generic(q, bio, rw, BLK_TA_GETRQ);
2153 out:
2154 return rq;
2158 * No available requests for this queue, unplug the device and wait for some
2159 * requests to become available.
2161 * Called with q->queue_lock held, and returns with it unlocked.
2163 static struct request *get_request_wait(request_queue_t *q, int rw_flags,
2164 struct bio *bio)
2166 const int rw = rw_flags & 0x01;
2167 struct request *rq;
2169 rq = get_request(q, rw_flags, bio, GFP_NOIO);
2170 while (!rq) {
2171 DEFINE_WAIT(wait);
2172 struct request_list *rl = &q->rq;
2174 prepare_to_wait_exclusive(&rl->wait[rw], &wait,
2175 TASK_UNINTERRUPTIBLE);
2177 rq = get_request(q, rw_flags, bio, GFP_NOIO);
2179 if (!rq) {
2180 struct io_context *ioc;
2182 blk_add_trace_generic(q, bio, rw, BLK_TA_SLEEPRQ);
2184 __generic_unplug_device(q);
2185 spin_unlock_irq(q->queue_lock);
2186 io_schedule();
2189 * After sleeping, we become a "batching" process and
2190 * will be able to allocate at least one request, and
2191 * up to a big batch of them for a small period time.
2192 * See ioc_batching, ioc_set_batching
2194 ioc = current_io_context(GFP_NOIO, q->node);
2195 ioc_set_batching(q, ioc);
2197 spin_lock_irq(q->queue_lock);
2199 finish_wait(&rl->wait[rw], &wait);
2202 return rq;
2205 struct request *blk_get_request(request_queue_t *q, int rw, gfp_t gfp_mask)
2207 struct request *rq;
2209 BUG_ON(rw != READ && rw != WRITE);
2211 spin_lock_irq(q->queue_lock);
2212 if (gfp_mask & __GFP_WAIT) {
2213 rq = get_request_wait(q, rw, NULL);
2214 } else {
2215 rq = get_request(q, rw, NULL, gfp_mask);
2216 if (!rq)
2217 spin_unlock_irq(q->queue_lock);
2219 /* q->queue_lock is unlocked at this point */
2221 return rq;
2223 EXPORT_SYMBOL(blk_get_request);
2226 * blk_start_queueing - initiate dispatch of requests to device
2227 * @q: request queue to kick into gear
2229 * This is basically a helper to remove the need to know whether a queue
2230 * is plugged or not if someone just wants to initiate dispatch of requests
2231 * for this queue.
2233 * The queue lock must be held with interrupts disabled.
2235 void blk_start_queueing(request_queue_t *q)
2237 if (!blk_queue_plugged(q))
2238 q->request_fn(q);
2239 else
2240 __generic_unplug_device(q);
2242 EXPORT_SYMBOL(blk_start_queueing);
2245 * blk_requeue_request - put a request back on queue
2246 * @q: request queue where request should be inserted
2247 * @rq: request to be inserted
2249 * Description:
2250 * Drivers often keep queueing requests until the hardware cannot accept
2251 * more, when that condition happens we need to put the request back
2252 * on the queue. Must be called with queue lock held.
2254 void blk_requeue_request(request_queue_t *q, struct request *rq)
2256 blk_add_trace_rq(q, rq, BLK_TA_REQUEUE);
2258 if (blk_rq_tagged(rq))
2259 blk_queue_end_tag(q, rq);
2261 elv_requeue_request(q, rq);
2264 EXPORT_SYMBOL(blk_requeue_request);
2267 * blk_insert_request - insert a special request in to a request queue
2268 * @q: request queue where request should be inserted
2269 * @rq: request to be inserted
2270 * @at_head: insert request at head or tail of queue
2271 * @data: private data
2273 * Description:
2274 * Many block devices need to execute commands asynchronously, so they don't
2275 * block the whole kernel from preemption during request execution. This is
2276 * accomplished normally by inserting aritficial requests tagged as
2277 * REQ_SPECIAL in to the corresponding request queue, and letting them be
2278 * scheduled for actual execution by the request queue.
2280 * We have the option of inserting the head or the tail of the queue.
2281 * Typically we use the tail for new ioctls and so forth. We use the head
2282 * of the queue for things like a QUEUE_FULL message from a device, or a
2283 * host that is unable to accept a particular command.
2285 void blk_insert_request(request_queue_t *q, struct request *rq,
2286 int at_head, void *data)
2288 int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2289 unsigned long flags;
2292 * tell I/O scheduler that this isn't a regular read/write (ie it
2293 * must not attempt merges on this) and that it acts as a soft
2294 * barrier
2296 rq->cmd_type = REQ_TYPE_SPECIAL;
2297 rq->cmd_flags |= REQ_SOFTBARRIER;
2299 rq->special = data;
2301 spin_lock_irqsave(q->queue_lock, flags);
2304 * If command is tagged, release the tag
2306 if (blk_rq_tagged(rq))
2307 blk_queue_end_tag(q, rq);
2309 drive_stat_acct(rq, rq->nr_sectors, 1);
2310 __elv_add_request(q, rq, where, 0);
2311 blk_start_queueing(q);
2312 spin_unlock_irqrestore(q->queue_lock, flags);
2315 EXPORT_SYMBOL(blk_insert_request);
2317 static int __blk_rq_unmap_user(struct bio *bio)
2319 int ret = 0;
2321 if (bio) {
2322 if (bio_flagged(bio, BIO_USER_MAPPED))
2323 bio_unmap_user(bio);
2324 else
2325 ret = bio_uncopy_user(bio);
2328 return ret;
2331 static int __blk_rq_map_user(request_queue_t *q, struct request *rq,
2332 void __user *ubuf, unsigned int len)
2334 unsigned long uaddr;
2335 struct bio *bio, *orig_bio;
2336 int reading, ret;
2338 reading = rq_data_dir(rq) == READ;
2341 * if alignment requirement is satisfied, map in user pages for
2342 * direct dma. else, set up kernel bounce buffers
2344 uaddr = (unsigned long) ubuf;
2345 if (!(uaddr & queue_dma_alignment(q)) && !(len & queue_dma_alignment(q)))
2346 bio = bio_map_user(q, NULL, uaddr, len, reading);
2347 else
2348 bio = bio_copy_user(q, uaddr, len, reading);
2350 if (IS_ERR(bio))
2351 return PTR_ERR(bio);
2353 orig_bio = bio;
2354 blk_queue_bounce(q, &bio);
2357 * We link the bounce buffer in and could have to traverse it
2358 * later so we have to get a ref to prevent it from being freed
2360 bio_get(bio);
2362 if (!rq->bio)
2363 blk_rq_bio_prep(q, rq, bio);
2364 else if (!ll_back_merge_fn(q, rq, bio)) {
2365 ret = -EINVAL;
2366 goto unmap_bio;
2367 } else {
2368 rq->biotail->bi_next = bio;
2369 rq->biotail = bio;
2371 rq->data_len += bio->bi_size;
2374 return bio->bi_size;
2376 unmap_bio:
2377 /* if it was boucned we must call the end io function */
2378 bio_endio(bio, bio->bi_size, 0);
2379 __blk_rq_unmap_user(orig_bio);
2380 bio_put(bio);
2381 return ret;
2385 * blk_rq_map_user - map user data to a request, for REQ_BLOCK_PC usage
2386 * @q: request queue where request should be inserted
2387 * @rq: request structure to fill
2388 * @ubuf: the user buffer
2389 * @len: length of user data
2391 * Description:
2392 * Data will be mapped directly for zero copy io, if possible. Otherwise
2393 * a kernel bounce buffer is used.
2395 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2396 * still in process context.
2398 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2399 * before being submitted to the device, as pages mapped may be out of
2400 * reach. It's the callers responsibility to make sure this happens. The
2401 * original bio must be passed back in to blk_rq_unmap_user() for proper
2402 * unmapping.
2404 int blk_rq_map_user(request_queue_t *q, struct request *rq, void __user *ubuf,
2405 unsigned long len)
2407 unsigned long bytes_read = 0;
2408 struct bio *bio = NULL;
2409 int ret;
2411 if (len > (q->max_hw_sectors << 9))
2412 return -EINVAL;
2413 if (!len || !ubuf)
2414 return -EINVAL;
2416 while (bytes_read != len) {
2417 unsigned long map_len, end, start;
2419 map_len = min_t(unsigned long, len - bytes_read, BIO_MAX_SIZE);
2420 end = ((unsigned long)ubuf + map_len + PAGE_SIZE - 1)
2421 >> PAGE_SHIFT;
2422 start = (unsigned long)ubuf >> PAGE_SHIFT;
2425 * A bad offset could cause us to require BIO_MAX_PAGES + 1
2426 * pages. If this happens we just lower the requested
2427 * mapping len by a page so that we can fit
2429 if (end - start > BIO_MAX_PAGES)
2430 map_len -= PAGE_SIZE;
2432 ret = __blk_rq_map_user(q, rq, ubuf, map_len);
2433 if (ret < 0)
2434 goto unmap_rq;
2435 if (!bio)
2436 bio = rq->bio;
2437 bytes_read += ret;
2438 ubuf += ret;
2441 rq->buffer = rq->data = NULL;
2442 return 0;
2443 unmap_rq:
2444 blk_rq_unmap_user(bio);
2445 return ret;
2448 EXPORT_SYMBOL(blk_rq_map_user);
2451 * blk_rq_map_user_iov - map user data to a request, for REQ_BLOCK_PC usage
2452 * @q: request queue where request should be inserted
2453 * @rq: request to map data to
2454 * @iov: pointer to the iovec
2455 * @iov_count: number of elements in the iovec
2456 * @len: I/O byte count
2458 * Description:
2459 * Data will be mapped directly for zero copy io, if possible. Otherwise
2460 * a kernel bounce buffer is used.
2462 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2463 * still in process context.
2465 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2466 * before being submitted to the device, as pages mapped may be out of
2467 * reach. It's the callers responsibility to make sure this happens. The
2468 * original bio must be passed back in to blk_rq_unmap_user() for proper
2469 * unmapping.
2471 int blk_rq_map_user_iov(request_queue_t *q, struct request *rq,
2472 struct sg_iovec *iov, int iov_count, unsigned int len)
2474 struct bio *bio;
2476 if (!iov || iov_count <= 0)
2477 return -EINVAL;
2479 /* we don't allow misaligned data like bio_map_user() does. If the
2480 * user is using sg, they're expected to know the alignment constraints
2481 * and respect them accordingly */
2482 bio = bio_map_user_iov(q, NULL, iov, iov_count, rq_data_dir(rq)== READ);
2483 if (IS_ERR(bio))
2484 return PTR_ERR(bio);
2486 if (bio->bi_size != len) {
2487 bio_endio(bio, bio->bi_size, 0);
2488 bio_unmap_user(bio);
2489 return -EINVAL;
2492 bio_get(bio);
2493 blk_rq_bio_prep(q, rq, bio);
2494 rq->buffer = rq->data = NULL;
2495 return 0;
2498 EXPORT_SYMBOL(blk_rq_map_user_iov);
2501 * blk_rq_unmap_user - unmap a request with user data
2502 * @bio: start of bio list
2504 * Description:
2505 * Unmap a rq previously mapped by blk_rq_map_user(). The caller must
2506 * supply the original rq->bio from the blk_rq_map_user() return, since
2507 * the io completion may have changed rq->bio.
2509 int blk_rq_unmap_user(struct bio *bio)
2511 struct bio *mapped_bio;
2512 int ret = 0, ret2;
2514 while (bio) {
2515 mapped_bio = bio;
2516 if (unlikely(bio_flagged(bio, BIO_BOUNCED)))
2517 mapped_bio = bio->bi_private;
2519 ret2 = __blk_rq_unmap_user(mapped_bio);
2520 if (ret2 && !ret)
2521 ret = ret2;
2523 mapped_bio = bio;
2524 bio = bio->bi_next;
2525 bio_put(mapped_bio);
2528 return ret;
2531 EXPORT_SYMBOL(blk_rq_unmap_user);
2534 * blk_rq_map_kern - map kernel data to a request, for REQ_BLOCK_PC usage
2535 * @q: request queue where request should be inserted
2536 * @rq: request to fill
2537 * @kbuf: the kernel buffer
2538 * @len: length of user data
2539 * @gfp_mask: memory allocation flags
2541 int blk_rq_map_kern(request_queue_t *q, struct request *rq, void *kbuf,
2542 unsigned int len, gfp_t gfp_mask)
2544 struct bio *bio;
2546 if (len > (q->max_hw_sectors << 9))
2547 return -EINVAL;
2548 if (!len || !kbuf)
2549 return -EINVAL;
2551 bio = bio_map_kern(q, kbuf, len, gfp_mask);
2552 if (IS_ERR(bio))
2553 return PTR_ERR(bio);
2555 if (rq_data_dir(rq) == WRITE)
2556 bio->bi_rw |= (1 << BIO_RW);
2558 blk_rq_bio_prep(q, rq, bio);
2559 rq->buffer = rq->data = NULL;
2560 return 0;
2563 EXPORT_SYMBOL(blk_rq_map_kern);
2566 * blk_execute_rq_nowait - insert a request into queue for execution
2567 * @q: queue to insert the request in
2568 * @bd_disk: matching gendisk
2569 * @rq: request to insert
2570 * @at_head: insert request at head or tail of queue
2571 * @done: I/O completion handler
2573 * Description:
2574 * Insert a fully prepared request at the back of the io scheduler queue
2575 * for execution. Don't wait for completion.
2577 void blk_execute_rq_nowait(request_queue_t *q, struct gendisk *bd_disk,
2578 struct request *rq, int at_head,
2579 rq_end_io_fn *done)
2581 int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2583 rq->rq_disk = bd_disk;
2584 rq->cmd_flags |= REQ_NOMERGE;
2585 rq->end_io = done;
2586 WARN_ON(irqs_disabled());
2587 spin_lock_irq(q->queue_lock);
2588 __elv_add_request(q, rq, where, 1);
2589 __generic_unplug_device(q);
2590 spin_unlock_irq(q->queue_lock);
2592 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
2595 * blk_execute_rq - insert a request into queue for execution
2596 * @q: queue to insert the request in
2597 * @bd_disk: matching gendisk
2598 * @rq: request to insert
2599 * @at_head: insert request at head or tail of queue
2601 * Description:
2602 * Insert a fully prepared request at the back of the io scheduler queue
2603 * for execution and wait for completion.
2605 int blk_execute_rq(request_queue_t *q, struct gendisk *bd_disk,
2606 struct request *rq, int at_head)
2608 DECLARE_COMPLETION_ONSTACK(wait);
2609 char sense[SCSI_SENSE_BUFFERSIZE];
2610 int err = 0;
2613 * we need an extra reference to the request, so we can look at
2614 * it after io completion
2616 rq->ref_count++;
2618 if (!rq->sense) {
2619 memset(sense, 0, sizeof(sense));
2620 rq->sense = sense;
2621 rq->sense_len = 0;
2624 rq->end_io_data = &wait;
2625 blk_execute_rq_nowait(q, bd_disk, rq, at_head, blk_end_sync_rq);
2626 wait_for_completion(&wait);
2628 if (rq->errors)
2629 err = -EIO;
2631 return err;
2634 EXPORT_SYMBOL(blk_execute_rq);
2637 * blkdev_issue_flush - queue a flush
2638 * @bdev: blockdev to issue flush for
2639 * @error_sector: error sector
2641 * Description:
2642 * Issue a flush for the block device in question. Caller can supply
2643 * room for storing the error offset in case of a flush error, if they
2644 * wish to. Caller must run wait_for_completion() on its own.
2646 int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
2648 request_queue_t *q;
2650 if (bdev->bd_disk == NULL)
2651 return -ENXIO;
2653 q = bdev_get_queue(bdev);
2654 if (!q)
2655 return -ENXIO;
2656 if (!q->issue_flush_fn)
2657 return -EOPNOTSUPP;
2659 return q->issue_flush_fn(q, bdev->bd_disk, error_sector);
2662 EXPORT_SYMBOL(blkdev_issue_flush);
2664 static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io)
2666 int rw = rq_data_dir(rq);
2668 if (!blk_fs_request(rq) || !rq->rq_disk)
2669 return;
2671 if (!new_io) {
2672 __disk_stat_inc(rq->rq_disk, merges[rw]);
2673 } else {
2674 disk_round_stats(rq->rq_disk);
2675 rq->rq_disk->in_flight++;
2680 * add-request adds a request to the linked list.
2681 * queue lock is held and interrupts disabled, as we muck with the
2682 * request queue list.
2684 static inline void add_request(request_queue_t * q, struct request * req)
2686 drive_stat_acct(req, req->nr_sectors, 1);
2689 * elevator indicated where it wants this request to be
2690 * inserted at elevator_merge time
2692 __elv_add_request(q, req, ELEVATOR_INSERT_SORT, 0);
2696 * disk_round_stats() - Round off the performance stats on a struct
2697 * disk_stats.
2699 * The average IO queue length and utilisation statistics are maintained
2700 * by observing the current state of the queue length and the amount of
2701 * time it has been in this state for.
2703 * Normally, that accounting is done on IO completion, but that can result
2704 * in more than a second's worth of IO being accounted for within any one
2705 * second, leading to >100% utilisation. To deal with that, we call this
2706 * function to do a round-off before returning the results when reading
2707 * /proc/diskstats. This accounts immediately for all queue usage up to
2708 * the current jiffies and restarts the counters again.
2710 void disk_round_stats(struct gendisk *disk)
2712 unsigned long now = jiffies;
2714 if (now == disk->stamp)
2715 return;
2717 if (disk->in_flight) {
2718 __disk_stat_add(disk, time_in_queue,
2719 disk->in_flight * (now - disk->stamp));
2720 __disk_stat_add(disk, io_ticks, (now - disk->stamp));
2722 disk->stamp = now;
2725 EXPORT_SYMBOL_GPL(disk_round_stats);
2728 * queue lock must be held
2730 void __blk_put_request(request_queue_t *q, struct request *req)
2732 if (unlikely(!q))
2733 return;
2734 if (unlikely(--req->ref_count))
2735 return;
2737 elv_completed_request(q, req);
2740 * Request may not have originated from ll_rw_blk. if not,
2741 * it didn't come out of our reserved rq pools
2743 if (req->cmd_flags & REQ_ALLOCED) {
2744 int rw = rq_data_dir(req);
2745 int priv = req->cmd_flags & REQ_ELVPRIV;
2747 BUG_ON(!list_empty(&req->queuelist));
2748 BUG_ON(!hlist_unhashed(&req->hash));
2750 blk_free_request(q, req);
2751 freed_request(q, rw, priv);
2755 EXPORT_SYMBOL_GPL(__blk_put_request);
2757 void blk_put_request(struct request *req)
2759 unsigned long flags;
2760 request_queue_t *q = req->q;
2763 * Gee, IDE calls in w/ NULL q. Fix IDE and remove the
2764 * following if (q) test.
2766 if (q) {
2767 spin_lock_irqsave(q->queue_lock, flags);
2768 __blk_put_request(q, req);
2769 spin_unlock_irqrestore(q->queue_lock, flags);
2773 EXPORT_SYMBOL(blk_put_request);
2776 * blk_end_sync_rq - executes a completion event on a request
2777 * @rq: request to complete
2778 * @error: end io status of the request
2780 void blk_end_sync_rq(struct request *rq, int error)
2782 struct completion *waiting = rq->end_io_data;
2784 rq->end_io_data = NULL;
2785 __blk_put_request(rq->q, rq);
2788 * complete last, if this is a stack request the process (and thus
2789 * the rq pointer) could be invalid right after this complete()
2791 complete(waiting);
2793 EXPORT_SYMBOL(blk_end_sync_rq);
2796 * Has to be called with the request spinlock acquired
2798 static int attempt_merge(request_queue_t *q, struct request *req,
2799 struct request *next)
2801 if (!rq_mergeable(req) || !rq_mergeable(next))
2802 return 0;
2805 * not contiguous
2807 if (req->sector + req->nr_sectors != next->sector)
2808 return 0;
2810 if (rq_data_dir(req) != rq_data_dir(next)
2811 || req->rq_disk != next->rq_disk
2812 || next->special)
2813 return 0;
2816 * If we are allowed to merge, then append bio list
2817 * from next to rq and release next. merge_requests_fn
2818 * will have updated segment counts, update sector
2819 * counts here.
2821 if (!ll_merge_requests_fn(q, req, next))
2822 return 0;
2825 * At this point we have either done a back merge
2826 * or front merge. We need the smaller start_time of
2827 * the merged requests to be the current request
2828 * for accounting purposes.
2830 if (time_after(req->start_time, next->start_time))
2831 req->start_time = next->start_time;
2833 req->biotail->bi_next = next->bio;
2834 req->biotail = next->biotail;
2836 req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
2838 elv_merge_requests(q, req, next);
2840 if (req->rq_disk) {
2841 disk_round_stats(req->rq_disk);
2842 req->rq_disk->in_flight--;
2845 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
2847 __blk_put_request(q, next);
2848 return 1;
2851 static inline int attempt_back_merge(request_queue_t *q, struct request *rq)
2853 struct request *next = elv_latter_request(q, rq);
2855 if (next)
2856 return attempt_merge(q, rq, next);
2858 return 0;
2861 static inline int attempt_front_merge(request_queue_t *q, struct request *rq)
2863 struct request *prev = elv_former_request(q, rq);
2865 if (prev)
2866 return attempt_merge(q, prev, rq);
2868 return 0;
2871 static void init_request_from_bio(struct request *req, struct bio *bio)
2873 req->cmd_type = REQ_TYPE_FS;
2876 * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST)
2878 if (bio_rw_ahead(bio) || bio_failfast(bio))
2879 req->cmd_flags |= REQ_FAILFAST;
2882 * REQ_BARRIER implies no merging, but lets make it explicit
2884 if (unlikely(bio_barrier(bio)))
2885 req->cmd_flags |= (REQ_HARDBARRIER | REQ_NOMERGE);
2887 if (bio_sync(bio))
2888 req->cmd_flags |= REQ_RW_SYNC;
2889 if (bio_rw_meta(bio))
2890 req->cmd_flags |= REQ_RW_META;
2892 req->errors = 0;
2893 req->hard_sector = req->sector = bio->bi_sector;
2894 req->hard_nr_sectors = req->nr_sectors = bio_sectors(bio);
2895 req->current_nr_sectors = req->hard_cur_sectors = bio_cur_sectors(bio);
2896 req->nr_phys_segments = bio_phys_segments(req->q, bio);
2897 req->nr_hw_segments = bio_hw_segments(req->q, bio);
2898 req->buffer = bio_data(bio); /* see ->buffer comment above */
2899 req->bio = req->biotail = bio;
2900 req->ioprio = bio_prio(bio);
2901 req->rq_disk = bio->bi_bdev->bd_disk;
2902 req->start_time = jiffies;
2905 static int __make_request(request_queue_t *q, struct bio *bio)
2907 struct request *req;
2908 int el_ret, nr_sectors, barrier, err;
2909 const unsigned short prio = bio_prio(bio);
2910 const int sync = bio_sync(bio);
2911 int rw_flags;
2913 nr_sectors = bio_sectors(bio);
2916 * low level driver can indicate that it wants pages above a
2917 * certain limit bounced to low memory (ie for highmem, or even
2918 * ISA dma in theory)
2920 blk_queue_bounce(q, &bio);
2922 barrier = bio_barrier(bio);
2923 if (unlikely(barrier) && (q->next_ordered == QUEUE_ORDERED_NONE)) {
2924 err = -EOPNOTSUPP;
2925 goto end_io;
2928 spin_lock_irq(q->queue_lock);
2930 if (unlikely(barrier) || elv_queue_empty(q))
2931 goto get_rq;
2933 el_ret = elv_merge(q, &req, bio);
2934 switch (el_ret) {
2935 case ELEVATOR_BACK_MERGE:
2936 BUG_ON(!rq_mergeable(req));
2938 if (!ll_back_merge_fn(q, req, bio))
2939 break;
2941 blk_add_trace_bio(q, bio, BLK_TA_BACKMERGE);
2943 req->biotail->bi_next = bio;
2944 req->biotail = bio;
2945 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2946 req->ioprio = ioprio_best(req->ioprio, prio);
2947 drive_stat_acct(req, nr_sectors, 0);
2948 if (!attempt_back_merge(q, req))
2949 elv_merged_request(q, req, el_ret);
2950 goto out;
2952 case ELEVATOR_FRONT_MERGE:
2953 BUG_ON(!rq_mergeable(req));
2955 if (!ll_front_merge_fn(q, req, bio))
2956 break;
2958 blk_add_trace_bio(q, bio, BLK_TA_FRONTMERGE);
2960 bio->bi_next = req->bio;
2961 req->bio = bio;
2964 * may not be valid. if the low level driver said
2965 * it didn't need a bounce buffer then it better
2966 * not touch req->buffer either...
2968 req->buffer = bio_data(bio);
2969 req->current_nr_sectors = bio_cur_sectors(bio);
2970 req->hard_cur_sectors = req->current_nr_sectors;
2971 req->sector = req->hard_sector = bio->bi_sector;
2972 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2973 req->ioprio = ioprio_best(req->ioprio, prio);
2974 drive_stat_acct(req, nr_sectors, 0);
2975 if (!attempt_front_merge(q, req))
2976 elv_merged_request(q, req, el_ret);
2977 goto out;
2979 /* ELV_NO_MERGE: elevator says don't/can't merge. */
2980 default:
2984 get_rq:
2986 * This sync check and mask will be re-done in init_request_from_bio(),
2987 * but we need to set it earlier to expose the sync flag to the
2988 * rq allocator and io schedulers.
2990 rw_flags = bio_data_dir(bio);
2991 if (sync)
2992 rw_flags |= REQ_RW_SYNC;
2995 * Grab a free request. This is might sleep but can not fail.
2996 * Returns with the queue unlocked.
2998 req = get_request_wait(q, rw_flags, bio);
3001 * After dropping the lock and possibly sleeping here, our request
3002 * may now be mergeable after it had proven unmergeable (above).
3003 * We don't worry about that case for efficiency. It won't happen
3004 * often, and the elevators are able to handle it.
3006 init_request_from_bio(req, bio);
3008 spin_lock_irq(q->queue_lock);
3009 if (elv_queue_empty(q))
3010 blk_plug_device(q);
3011 add_request(q, req);
3012 out:
3013 if (sync)
3014 __generic_unplug_device(q);
3016 spin_unlock_irq(q->queue_lock);
3017 return 0;
3019 end_io:
3020 bio_endio(bio, nr_sectors << 9, err);
3021 return 0;
3025 * If bio->bi_dev is a partition, remap the location
3027 static inline void blk_partition_remap(struct bio *bio)
3029 struct block_device *bdev = bio->bi_bdev;
3031 if (bdev != bdev->bd_contains) {
3032 struct hd_struct *p = bdev->bd_part;
3033 const int rw = bio_data_dir(bio);
3035 p->sectors[rw] += bio_sectors(bio);
3036 p->ios[rw]++;
3038 bio->bi_sector += p->start_sect;
3039 bio->bi_bdev = bdev->bd_contains;
3043 static void handle_bad_sector(struct bio *bio)
3045 char b[BDEVNAME_SIZE];
3047 printk(KERN_INFO "attempt to access beyond end of device\n");
3048 printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
3049 bdevname(bio->bi_bdev, b),
3050 bio->bi_rw,
3051 (unsigned long long)bio->bi_sector + bio_sectors(bio),
3052 (long long)(bio->bi_bdev->bd_inode->i_size >> 9));
3054 set_bit(BIO_EOF, &bio->bi_flags);
3057 #ifdef CONFIG_FAIL_MAKE_REQUEST
3059 static DECLARE_FAULT_ATTR(fail_make_request);
3061 static int __init setup_fail_make_request(char *str)
3063 return setup_fault_attr(&fail_make_request, str);
3065 __setup("fail_make_request=", setup_fail_make_request);
3067 static int should_fail_request(struct bio *bio)
3069 if ((bio->bi_bdev->bd_disk->flags & GENHD_FL_FAIL) ||
3070 (bio->bi_bdev->bd_part && bio->bi_bdev->bd_part->make_it_fail))
3071 return should_fail(&fail_make_request, bio->bi_size);
3073 return 0;
3076 static int __init fail_make_request_debugfs(void)
3078 return init_fault_attr_dentries(&fail_make_request,
3079 "fail_make_request");
3082 late_initcall(fail_make_request_debugfs);
3084 #else /* CONFIG_FAIL_MAKE_REQUEST */
3086 static inline int should_fail_request(struct bio *bio)
3088 return 0;
3091 #endif /* CONFIG_FAIL_MAKE_REQUEST */
3094 * generic_make_request: hand a buffer to its device driver for I/O
3095 * @bio: The bio describing the location in memory and on the device.
3097 * generic_make_request() is used to make I/O requests of block
3098 * devices. It is passed a &struct bio, which describes the I/O that needs
3099 * to be done.
3101 * generic_make_request() does not return any status. The
3102 * success/failure status of the request, along with notification of
3103 * completion, is delivered asynchronously through the bio->bi_end_io
3104 * function described (one day) else where.
3106 * The caller of generic_make_request must make sure that bi_io_vec
3107 * are set to describe the memory buffer, and that bi_dev and bi_sector are
3108 * set to describe the device address, and the
3109 * bi_end_io and optionally bi_private are set to describe how
3110 * completion notification should be signaled.
3112 * generic_make_request and the drivers it calls may use bi_next if this
3113 * bio happens to be merged with someone else, and may change bi_dev and
3114 * bi_sector for remaps as it sees fit. So the values of these fields
3115 * should NOT be depended on after the call to generic_make_request.
3117 void generic_make_request(struct bio *bio)
3119 request_queue_t *q;
3120 sector_t maxsector;
3121 sector_t old_sector;
3122 int ret, nr_sectors = bio_sectors(bio);
3123 dev_t old_dev;
3125 might_sleep();
3126 /* Test device or partition size, when known. */
3127 maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
3128 if (maxsector) {
3129 sector_t sector = bio->bi_sector;
3131 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
3133 * This may well happen - the kernel calls bread()
3134 * without checking the size of the device, e.g., when
3135 * mounting a device.
3137 handle_bad_sector(bio);
3138 goto end_io;
3143 * Resolve the mapping until finished. (drivers are
3144 * still free to implement/resolve their own stacking
3145 * by explicitly returning 0)
3147 * NOTE: we don't repeat the blk_size check for each new device.
3148 * Stacking drivers are expected to know what they are doing.
3150 old_sector = -1;
3151 old_dev = 0;
3152 do {
3153 char b[BDEVNAME_SIZE];
3155 q = bdev_get_queue(bio->bi_bdev);
3156 if (!q) {
3157 printk(KERN_ERR
3158 "generic_make_request: Trying to access "
3159 "nonexistent block-device %s (%Lu)\n",
3160 bdevname(bio->bi_bdev, b),
3161 (long long) bio->bi_sector);
3162 end_io:
3163 bio_endio(bio, bio->bi_size, -EIO);
3164 break;
3167 if (unlikely(bio_sectors(bio) > q->max_hw_sectors)) {
3168 printk("bio too big device %s (%u > %u)\n",
3169 bdevname(bio->bi_bdev, b),
3170 bio_sectors(bio),
3171 q->max_hw_sectors);
3172 goto end_io;
3175 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)))
3176 goto end_io;
3178 if (should_fail_request(bio))
3179 goto end_io;
3182 * If this device has partitions, remap block n
3183 * of partition p to block n+start(p) of the disk.
3185 blk_partition_remap(bio);
3187 if (old_sector != -1)
3188 blk_add_trace_remap(q, bio, old_dev, bio->bi_sector,
3189 old_sector);
3191 blk_add_trace_bio(q, bio, BLK_TA_QUEUE);
3193 old_sector = bio->bi_sector;
3194 old_dev = bio->bi_bdev->bd_dev;
3196 maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
3197 if (maxsector) {
3198 sector_t sector = bio->bi_sector;
3200 if (maxsector < nr_sectors ||
3201 maxsector - nr_sectors < sector) {
3203 * This may well happen - partitions are not
3204 * checked to make sure they are within the size
3205 * of the whole device.
3207 handle_bad_sector(bio);
3208 goto end_io;
3212 ret = q->make_request_fn(q, bio);
3213 } while (ret);
3216 EXPORT_SYMBOL(generic_make_request);
3219 * submit_bio: submit a bio to the block device layer for I/O
3220 * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
3221 * @bio: The &struct bio which describes the I/O
3223 * submit_bio() is very similar in purpose to generic_make_request(), and
3224 * uses that function to do most of the work. Both are fairly rough
3225 * interfaces, @bio must be presetup and ready for I/O.
3228 void submit_bio(int rw, struct bio *bio)
3230 int count = bio_sectors(bio);
3232 BIO_BUG_ON(!bio->bi_size);
3233 BIO_BUG_ON(!bio->bi_io_vec);
3234 bio->bi_rw |= rw;
3235 if (rw & WRITE) {
3236 count_vm_events(PGPGOUT, count);
3237 } else {
3238 task_io_account_read(bio->bi_size);
3239 count_vm_events(PGPGIN, count);
3242 if (unlikely(block_dump)) {
3243 char b[BDEVNAME_SIZE];
3244 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s\n",
3245 current->comm, current->pid,
3246 (rw & WRITE) ? "WRITE" : "READ",
3247 (unsigned long long)bio->bi_sector,
3248 bdevname(bio->bi_bdev,b));
3251 generic_make_request(bio);
3254 EXPORT_SYMBOL(submit_bio);
3256 static void blk_recalc_rq_segments(struct request *rq)
3258 struct bio *bio, *prevbio = NULL;
3259 int nr_phys_segs, nr_hw_segs;
3260 unsigned int phys_size, hw_size;
3261 request_queue_t *q = rq->q;
3263 if (!rq->bio)
3264 return;
3266 phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
3267 rq_for_each_bio(bio, rq) {
3268 /* Force bio hw/phys segs to be recalculated. */
3269 bio->bi_flags &= ~(1 << BIO_SEG_VALID);
3271 nr_phys_segs += bio_phys_segments(q, bio);
3272 nr_hw_segs += bio_hw_segments(q, bio);
3273 if (prevbio) {
3274 int pseg = phys_size + prevbio->bi_size + bio->bi_size;
3275 int hseg = hw_size + prevbio->bi_size + bio->bi_size;
3277 if (blk_phys_contig_segment(q, prevbio, bio) &&
3278 pseg <= q->max_segment_size) {
3279 nr_phys_segs--;
3280 phys_size += prevbio->bi_size + bio->bi_size;
3281 } else
3282 phys_size = 0;
3284 if (blk_hw_contig_segment(q, prevbio, bio) &&
3285 hseg <= q->max_segment_size) {
3286 nr_hw_segs--;
3287 hw_size += prevbio->bi_size + bio->bi_size;
3288 } else
3289 hw_size = 0;
3291 prevbio = bio;
3294 rq->nr_phys_segments = nr_phys_segs;
3295 rq->nr_hw_segments = nr_hw_segs;
3298 static void blk_recalc_rq_sectors(struct request *rq, int nsect)
3300 if (blk_fs_request(rq)) {
3301 rq->hard_sector += nsect;
3302 rq->hard_nr_sectors -= nsect;
3305 * Move the I/O submission pointers ahead if required.
3307 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
3308 (rq->sector <= rq->hard_sector)) {
3309 rq->sector = rq->hard_sector;
3310 rq->nr_sectors = rq->hard_nr_sectors;
3311 rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
3312 rq->current_nr_sectors = rq->hard_cur_sectors;
3313 rq->buffer = bio_data(rq->bio);
3317 * if total number of sectors is less than the first segment
3318 * size, something has gone terribly wrong
3320 if (rq->nr_sectors < rq->current_nr_sectors) {
3321 printk("blk: request botched\n");
3322 rq->nr_sectors = rq->current_nr_sectors;
3327 static int __end_that_request_first(struct request *req, int uptodate,
3328 int nr_bytes)
3330 int total_bytes, bio_nbytes, error, next_idx = 0;
3331 struct bio *bio;
3333 blk_add_trace_rq(req->q, req, BLK_TA_COMPLETE);
3336 * extend uptodate bool to allow < 0 value to be direct io error
3338 error = 0;
3339 if (end_io_error(uptodate))
3340 error = !uptodate ? -EIO : uptodate;
3343 * for a REQ_BLOCK_PC request, we want to carry any eventual
3344 * sense key with us all the way through
3346 if (!blk_pc_request(req))
3347 req->errors = 0;
3349 if (!uptodate) {
3350 if (blk_fs_request(req) && !(req->cmd_flags & REQ_QUIET))
3351 printk("end_request: I/O error, dev %s, sector %llu\n",
3352 req->rq_disk ? req->rq_disk->disk_name : "?",
3353 (unsigned long long)req->sector);
3356 if (blk_fs_request(req) && req->rq_disk) {
3357 const int rw = rq_data_dir(req);
3359 disk_stat_add(req->rq_disk, sectors[rw], nr_bytes >> 9);
3362 total_bytes = bio_nbytes = 0;
3363 while ((bio = req->bio) != NULL) {
3364 int nbytes;
3366 if (nr_bytes >= bio->bi_size) {
3367 req->bio = bio->bi_next;
3368 nbytes = bio->bi_size;
3369 if (!ordered_bio_endio(req, bio, nbytes, error))
3370 bio_endio(bio, nbytes, error);
3371 next_idx = 0;
3372 bio_nbytes = 0;
3373 } else {
3374 int idx = bio->bi_idx + next_idx;
3376 if (unlikely(bio->bi_idx >= bio->bi_vcnt)) {
3377 blk_dump_rq_flags(req, "__end_that");
3378 printk("%s: bio idx %d >= vcnt %d\n",
3379 __FUNCTION__,
3380 bio->bi_idx, bio->bi_vcnt);
3381 break;
3384 nbytes = bio_iovec_idx(bio, idx)->bv_len;
3385 BIO_BUG_ON(nbytes > bio->bi_size);
3388 * not a complete bvec done
3390 if (unlikely(nbytes > nr_bytes)) {
3391 bio_nbytes += nr_bytes;
3392 total_bytes += nr_bytes;
3393 break;
3397 * advance to the next vector
3399 next_idx++;
3400 bio_nbytes += nbytes;
3403 total_bytes += nbytes;
3404 nr_bytes -= nbytes;
3406 if ((bio = req->bio)) {
3408 * end more in this run, or just return 'not-done'
3410 if (unlikely(nr_bytes <= 0))
3411 break;
3416 * completely done
3418 if (!req->bio)
3419 return 0;
3422 * if the request wasn't completed, update state
3424 if (bio_nbytes) {
3425 if (!ordered_bio_endio(req, bio, bio_nbytes, error))
3426 bio_endio(bio, bio_nbytes, error);
3427 bio->bi_idx += next_idx;
3428 bio_iovec(bio)->bv_offset += nr_bytes;
3429 bio_iovec(bio)->bv_len -= nr_bytes;
3432 blk_recalc_rq_sectors(req, total_bytes >> 9);
3433 blk_recalc_rq_segments(req);
3434 return 1;
3438 * end_that_request_first - end I/O on a request
3439 * @req: the request being processed
3440 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3441 * @nr_sectors: number of sectors to end I/O on
3443 * Description:
3444 * Ends I/O on a number of sectors attached to @req, and sets it up
3445 * for the next range of segments (if any) in the cluster.
3447 * Return:
3448 * 0 - we are done with this request, call end_that_request_last()
3449 * 1 - still buffers pending for this request
3451 int end_that_request_first(struct request *req, int uptodate, int nr_sectors)
3453 return __end_that_request_first(req, uptodate, nr_sectors << 9);
3456 EXPORT_SYMBOL(end_that_request_first);
3459 * end_that_request_chunk - end I/O on a request
3460 * @req: the request being processed
3461 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3462 * @nr_bytes: number of bytes to complete
3464 * Description:
3465 * Ends I/O on a number of bytes attached to @req, and sets it up
3466 * for the next range of segments (if any). Like end_that_request_first(),
3467 * but deals with bytes instead of sectors.
3469 * Return:
3470 * 0 - we are done with this request, call end_that_request_last()
3471 * 1 - still buffers pending for this request
3473 int end_that_request_chunk(struct request *req, int uptodate, int nr_bytes)
3475 return __end_that_request_first(req, uptodate, nr_bytes);
3478 EXPORT_SYMBOL(end_that_request_chunk);
3481 * splice the completion data to a local structure and hand off to
3482 * process_completion_queue() to complete the requests
3484 static void blk_done_softirq(struct softirq_action *h)
3486 struct list_head *cpu_list, local_list;
3488 local_irq_disable();
3489 cpu_list = &__get_cpu_var(blk_cpu_done);
3490 list_replace_init(cpu_list, &local_list);
3491 local_irq_enable();
3493 while (!list_empty(&local_list)) {
3494 struct request *rq = list_entry(local_list.next, struct request, donelist);
3496 list_del_init(&rq->donelist);
3497 rq->q->softirq_done_fn(rq);
3501 static int blk_cpu_notify(struct notifier_block *self, unsigned long action,
3502 void *hcpu)
3505 * If a CPU goes away, splice its entries to the current CPU
3506 * and trigger a run of the softirq
3508 if (action == CPU_DEAD) {
3509 int cpu = (unsigned long) hcpu;
3511 local_irq_disable();
3512 list_splice_init(&per_cpu(blk_cpu_done, cpu),
3513 &__get_cpu_var(blk_cpu_done));
3514 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3515 local_irq_enable();
3518 return NOTIFY_OK;
3522 static struct notifier_block __devinitdata blk_cpu_notifier = {
3523 .notifier_call = blk_cpu_notify,
3527 * blk_complete_request - end I/O on a request
3528 * @req: the request being processed
3530 * Description:
3531 * Ends all I/O on a request. It does not handle partial completions,
3532 * unless the driver actually implements this in its completion callback
3533 * through requeueing. Theh actual completion happens out-of-order,
3534 * through a softirq handler. The user must have registered a completion
3535 * callback through blk_queue_softirq_done().
3538 void blk_complete_request(struct request *req)
3540 struct list_head *cpu_list;
3541 unsigned long flags;
3543 BUG_ON(!req->q->softirq_done_fn);
3545 local_irq_save(flags);
3547 cpu_list = &__get_cpu_var(blk_cpu_done);
3548 list_add_tail(&req->donelist, cpu_list);
3549 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3551 local_irq_restore(flags);
3554 EXPORT_SYMBOL(blk_complete_request);
3557 * queue lock must be held
3559 void end_that_request_last(struct request *req, int uptodate)
3561 struct gendisk *disk = req->rq_disk;
3562 int error;
3565 * extend uptodate bool to allow < 0 value to be direct io error
3567 error = 0;
3568 if (end_io_error(uptodate))
3569 error = !uptodate ? -EIO : uptodate;
3571 if (unlikely(laptop_mode) && blk_fs_request(req))
3572 laptop_io_completion();
3575 * Account IO completion. bar_rq isn't accounted as a normal
3576 * IO on queueing nor completion. Accounting the containing
3577 * request is enough.
3579 if (disk && blk_fs_request(req) && req != &req->q->bar_rq) {
3580 unsigned long duration = jiffies - req->start_time;
3581 const int rw = rq_data_dir(req);
3583 __disk_stat_inc(disk, ios[rw]);
3584 __disk_stat_add(disk, ticks[rw], duration);
3585 disk_round_stats(disk);
3586 disk->in_flight--;
3588 if (req->end_io)
3589 req->end_io(req, error);
3590 else
3591 __blk_put_request(req->q, req);
3594 EXPORT_SYMBOL(end_that_request_last);
3596 void end_request(struct request *req, int uptodate)
3598 if (!end_that_request_first(req, uptodate, req->hard_cur_sectors)) {
3599 add_disk_randomness(req->rq_disk);
3600 blkdev_dequeue_request(req);
3601 end_that_request_last(req, uptodate);
3605 EXPORT_SYMBOL(end_request);
3607 void blk_rq_bio_prep(request_queue_t *q, struct request *rq, struct bio *bio)
3609 /* first two bits are identical in rq->cmd_flags and bio->bi_rw */
3610 rq->cmd_flags |= (bio->bi_rw & 3);
3612 rq->nr_phys_segments = bio_phys_segments(q, bio);
3613 rq->nr_hw_segments = bio_hw_segments(q, bio);
3614 rq->current_nr_sectors = bio_cur_sectors(bio);
3615 rq->hard_cur_sectors = rq->current_nr_sectors;
3616 rq->hard_nr_sectors = rq->nr_sectors = bio_sectors(bio);
3617 rq->buffer = bio_data(bio);
3618 rq->data_len = bio->bi_size;
3620 rq->bio = rq->biotail = bio;
3623 EXPORT_SYMBOL(blk_rq_bio_prep);
3625 int kblockd_schedule_work(struct work_struct *work)
3627 return queue_work(kblockd_workqueue, work);
3630 EXPORT_SYMBOL(kblockd_schedule_work);
3632 void kblockd_flush(void)
3634 flush_workqueue(kblockd_workqueue);
3636 EXPORT_SYMBOL(kblockd_flush);
3638 int __init blk_dev_init(void)
3640 int i;
3642 kblockd_workqueue = create_workqueue("kblockd");
3643 if (!kblockd_workqueue)
3644 panic("Failed to create kblockd\n");
3646 request_cachep = kmem_cache_create("blkdev_requests",
3647 sizeof(struct request), 0, SLAB_PANIC, NULL, NULL);
3649 requestq_cachep = kmem_cache_create("blkdev_queue",
3650 sizeof(request_queue_t), 0, SLAB_PANIC, NULL, NULL);
3652 iocontext_cachep = kmem_cache_create("blkdev_ioc",
3653 sizeof(struct io_context), 0, SLAB_PANIC, NULL, NULL);
3655 for_each_possible_cpu(i)
3656 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3658 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq, NULL);
3659 register_hotcpu_notifier(&blk_cpu_notifier);
3661 blk_max_low_pfn = max_low_pfn - 1;
3662 blk_max_pfn = max_pfn - 1;
3664 return 0;
3668 * IO Context helper functions
3670 void put_io_context(struct io_context *ioc)
3672 if (ioc == NULL)
3673 return;
3675 BUG_ON(atomic_read(&ioc->refcount) == 0);
3677 if (atomic_dec_and_test(&ioc->refcount)) {
3678 struct cfq_io_context *cic;
3680 rcu_read_lock();
3681 if (ioc->aic && ioc->aic->dtor)
3682 ioc->aic->dtor(ioc->aic);
3683 if (ioc->cic_root.rb_node != NULL) {
3684 struct rb_node *n = rb_first(&ioc->cic_root);
3686 cic = rb_entry(n, struct cfq_io_context, rb_node);
3687 cic->dtor(ioc);
3689 rcu_read_unlock();
3691 kmem_cache_free(iocontext_cachep, ioc);
3694 EXPORT_SYMBOL(put_io_context);
3696 /* Called by the exitting task */
3697 void exit_io_context(void)
3699 struct io_context *ioc;
3700 struct cfq_io_context *cic;
3702 task_lock(current);
3703 ioc = current->io_context;
3704 current->io_context = NULL;
3705 task_unlock(current);
3707 ioc->task = NULL;
3708 if (ioc->aic && ioc->aic->exit)
3709 ioc->aic->exit(ioc->aic);
3710 if (ioc->cic_root.rb_node != NULL) {
3711 cic = rb_entry(rb_first(&ioc->cic_root), struct cfq_io_context, rb_node);
3712 cic->exit(ioc);
3715 put_io_context(ioc);
3719 * If the current task has no IO context then create one and initialise it.
3720 * Otherwise, return its existing IO context.
3722 * This returned IO context doesn't have a specifically elevated refcount,
3723 * but since the current task itself holds a reference, the context can be
3724 * used in general code, so long as it stays within `current` context.
3726 static struct io_context *current_io_context(gfp_t gfp_flags, int node)
3728 struct task_struct *tsk = current;
3729 struct io_context *ret;
3731 ret = tsk->io_context;
3732 if (likely(ret))
3733 return ret;
3735 ret = kmem_cache_alloc_node(iocontext_cachep, gfp_flags, node);
3736 if (ret) {
3737 atomic_set(&ret->refcount, 1);
3738 ret->task = current;
3739 ret->ioprio_changed = 0;
3740 ret->last_waited = jiffies; /* doesn't matter... */
3741 ret->nr_batch_requests = 0; /* because this is 0 */
3742 ret->aic = NULL;
3743 ret->cic_root.rb_node = NULL;
3744 /* make sure set_task_ioprio() sees the settings above */
3745 smp_wmb();
3746 tsk->io_context = ret;
3749 return ret;
3751 EXPORT_SYMBOL(current_io_context);
3754 * If the current task has no IO context then create one and initialise it.
3755 * If it does have a context, take a ref on it.
3757 * This is always called in the context of the task which submitted the I/O.
3759 struct io_context *get_io_context(gfp_t gfp_flags, int node)
3761 struct io_context *ret;
3762 ret = current_io_context(gfp_flags, node);
3763 if (likely(ret))
3764 atomic_inc(&ret->refcount);
3765 return ret;
3767 EXPORT_SYMBOL(get_io_context);
3769 void copy_io_context(struct io_context **pdst, struct io_context **psrc)
3771 struct io_context *src = *psrc;
3772 struct io_context *dst = *pdst;
3774 if (src) {
3775 BUG_ON(atomic_read(&src->refcount) == 0);
3776 atomic_inc(&src->refcount);
3777 put_io_context(dst);
3778 *pdst = src;
3781 EXPORT_SYMBOL(copy_io_context);
3783 void swap_io_context(struct io_context **ioc1, struct io_context **ioc2)
3785 struct io_context *temp;
3786 temp = *ioc1;
3787 *ioc1 = *ioc2;
3788 *ioc2 = temp;
3790 EXPORT_SYMBOL(swap_io_context);
3793 * sysfs parts below
3795 struct queue_sysfs_entry {
3796 struct attribute attr;
3797 ssize_t (*show)(struct request_queue *, char *);
3798 ssize_t (*store)(struct request_queue *, const char *, size_t);
3801 static ssize_t
3802 queue_var_show(unsigned int var, char *page)
3804 return sprintf(page, "%d\n", var);
3807 static ssize_t
3808 queue_var_store(unsigned long *var, const char *page, size_t count)
3810 char *p = (char *) page;
3812 *var = simple_strtoul(p, &p, 10);
3813 return count;
3816 static ssize_t queue_requests_show(struct request_queue *q, char *page)
3818 return queue_var_show(q->nr_requests, (page));
3821 static ssize_t
3822 queue_requests_store(struct request_queue *q, const char *page, size_t count)
3824 struct request_list *rl = &q->rq;
3825 unsigned long nr;
3826 int ret = queue_var_store(&nr, page, count);
3827 if (nr < BLKDEV_MIN_RQ)
3828 nr = BLKDEV_MIN_RQ;
3830 spin_lock_irq(q->queue_lock);
3831 q->nr_requests = nr;
3832 blk_queue_congestion_threshold(q);
3834 if (rl->count[READ] >= queue_congestion_on_threshold(q))
3835 blk_set_queue_congested(q, READ);
3836 else if (rl->count[READ] < queue_congestion_off_threshold(q))
3837 blk_clear_queue_congested(q, READ);
3839 if (rl->count[WRITE] >= queue_congestion_on_threshold(q))
3840 blk_set_queue_congested(q, WRITE);
3841 else if (rl->count[WRITE] < queue_congestion_off_threshold(q))
3842 blk_clear_queue_congested(q, WRITE);
3844 if (rl->count[READ] >= q->nr_requests) {
3845 blk_set_queue_full(q, READ);
3846 } else if (rl->count[READ]+1 <= q->nr_requests) {
3847 blk_clear_queue_full(q, READ);
3848 wake_up(&rl->wait[READ]);
3851 if (rl->count[WRITE] >= q->nr_requests) {
3852 blk_set_queue_full(q, WRITE);
3853 } else if (rl->count[WRITE]+1 <= q->nr_requests) {
3854 blk_clear_queue_full(q, WRITE);
3855 wake_up(&rl->wait[WRITE]);
3857 spin_unlock_irq(q->queue_lock);
3858 return ret;
3861 static ssize_t queue_ra_show(struct request_queue *q, char *page)
3863 int ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3865 return queue_var_show(ra_kb, (page));
3868 static ssize_t
3869 queue_ra_store(struct request_queue *q, const char *page, size_t count)
3871 unsigned long ra_kb;
3872 ssize_t ret = queue_var_store(&ra_kb, page, count);
3874 spin_lock_irq(q->queue_lock);
3875 q->backing_dev_info.ra_pages = ra_kb >> (PAGE_CACHE_SHIFT - 10);
3876 spin_unlock_irq(q->queue_lock);
3878 return ret;
3881 static ssize_t queue_max_sectors_show(struct request_queue *q, char *page)
3883 int max_sectors_kb = q->max_sectors >> 1;
3885 return queue_var_show(max_sectors_kb, (page));
3888 static ssize_t
3889 queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
3891 unsigned long max_sectors_kb,
3892 max_hw_sectors_kb = q->max_hw_sectors >> 1,
3893 page_kb = 1 << (PAGE_CACHE_SHIFT - 10);
3894 ssize_t ret = queue_var_store(&max_sectors_kb, page, count);
3895 int ra_kb;
3897 if (max_sectors_kb > max_hw_sectors_kb || max_sectors_kb < page_kb)
3898 return -EINVAL;
3900 * Take the queue lock to update the readahead and max_sectors
3901 * values synchronously:
3903 spin_lock_irq(q->queue_lock);
3905 * Trim readahead window as well, if necessary:
3907 ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3908 if (ra_kb > max_sectors_kb)
3909 q->backing_dev_info.ra_pages =
3910 max_sectors_kb >> (PAGE_CACHE_SHIFT - 10);
3912 q->max_sectors = max_sectors_kb << 1;
3913 spin_unlock_irq(q->queue_lock);
3915 return ret;
3918 static ssize_t queue_max_hw_sectors_show(struct request_queue *q, char *page)
3920 int max_hw_sectors_kb = q->max_hw_sectors >> 1;
3922 return queue_var_show(max_hw_sectors_kb, (page));
3926 static struct queue_sysfs_entry queue_requests_entry = {
3927 .attr = {.name = "nr_requests", .mode = S_IRUGO | S_IWUSR },
3928 .show = queue_requests_show,
3929 .store = queue_requests_store,
3932 static struct queue_sysfs_entry queue_ra_entry = {
3933 .attr = {.name = "read_ahead_kb", .mode = S_IRUGO | S_IWUSR },
3934 .show = queue_ra_show,
3935 .store = queue_ra_store,
3938 static struct queue_sysfs_entry queue_max_sectors_entry = {
3939 .attr = {.name = "max_sectors_kb", .mode = S_IRUGO | S_IWUSR },
3940 .show = queue_max_sectors_show,
3941 .store = queue_max_sectors_store,
3944 static struct queue_sysfs_entry queue_max_hw_sectors_entry = {
3945 .attr = {.name = "max_hw_sectors_kb", .mode = S_IRUGO },
3946 .show = queue_max_hw_sectors_show,
3949 static struct queue_sysfs_entry queue_iosched_entry = {
3950 .attr = {.name = "scheduler", .mode = S_IRUGO | S_IWUSR },
3951 .show = elv_iosched_show,
3952 .store = elv_iosched_store,
3955 static struct attribute *default_attrs[] = {
3956 &queue_requests_entry.attr,
3957 &queue_ra_entry.attr,
3958 &queue_max_hw_sectors_entry.attr,
3959 &queue_max_sectors_entry.attr,
3960 &queue_iosched_entry.attr,
3961 NULL,
3964 #define to_queue(atr) container_of((atr), struct queue_sysfs_entry, attr)
3966 static ssize_t
3967 queue_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
3969 struct queue_sysfs_entry *entry = to_queue(attr);
3970 request_queue_t *q = container_of(kobj, struct request_queue, kobj);
3971 ssize_t res;
3973 if (!entry->show)
3974 return -EIO;
3975 mutex_lock(&q->sysfs_lock);
3976 if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
3977 mutex_unlock(&q->sysfs_lock);
3978 return -ENOENT;
3980 res = entry->show(q, page);
3981 mutex_unlock(&q->sysfs_lock);
3982 return res;
3985 static ssize_t
3986 queue_attr_store(struct kobject *kobj, struct attribute *attr,
3987 const char *page, size_t length)
3989 struct queue_sysfs_entry *entry = to_queue(attr);
3990 request_queue_t *q = container_of(kobj, struct request_queue, kobj);
3992 ssize_t res;
3994 if (!entry->store)
3995 return -EIO;
3996 mutex_lock(&q->sysfs_lock);
3997 if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
3998 mutex_unlock(&q->sysfs_lock);
3999 return -ENOENT;
4001 res = entry->store(q, page, length);
4002 mutex_unlock(&q->sysfs_lock);
4003 return res;
4006 static struct sysfs_ops queue_sysfs_ops = {
4007 .show = queue_attr_show,
4008 .store = queue_attr_store,
4011 static struct kobj_type queue_ktype = {
4012 .sysfs_ops = &queue_sysfs_ops,
4013 .default_attrs = default_attrs,
4014 .release = blk_release_queue,
4017 int blk_register_queue(struct gendisk *disk)
4019 int ret;
4021 request_queue_t *q = disk->queue;
4023 if (!q || !q->request_fn)
4024 return -ENXIO;
4026 q->kobj.parent = kobject_get(&disk->kobj);
4028 ret = kobject_add(&q->kobj);
4029 if (ret < 0)
4030 return ret;
4032 kobject_uevent(&q->kobj, KOBJ_ADD);
4034 ret = elv_register_queue(q);
4035 if (ret) {
4036 kobject_uevent(&q->kobj, KOBJ_REMOVE);
4037 kobject_del(&q->kobj);
4038 return ret;
4041 return 0;
4044 void blk_unregister_queue(struct gendisk *disk)
4046 request_queue_t *q = disk->queue;
4048 if (q && q->request_fn) {
4049 elv_unregister_queue(q);
4051 kobject_uevent(&q->kobj, KOBJ_REMOVE);
4052 kobject_del(&q->kobj);
4053 kobject_put(&disk->kobj);